On the incrementality of pragmatic processing: An ERP investigation of informativeness and pragmatic abilities

Mante S Nieuwland; Tali Ditman; Gina R Kuperberg

doi:10.1016/j.jml.2010.06.005

. Author manuscript; available in PMC: 2011 Oct 1.

Published in final edited form as: J Mem Lang. 2010 Oct 1;63(3):324–346. doi: 10.1016/j.jml.2010.06.005

On the incrementality of pragmatic processing: An ERP investigation of informativeness and pragmatic abilities

Mante S Nieuwland ^1,^2,³, Tali Ditman ^2,³, Gina R Kuperberg ^2,^3,⁴

PMCID: PMC2950651 NIHMSID: NIHMS218221 PMID: 20936088

Abstract

In two event-related potential (ERP) experiments, we determined to what extent Grice’s maxim of informativeness as well as pragmatic ability contributes to the incremental build-up of sentence meaning, by examining the impact of underinformative versus informative scalar statements (e.g. “Some people have lungs/pets, and…”) on the N400 event-related potential (ERP), an electrophysiological index of semantic processing. In Experiment 1, only pragmatically skilled participants (as indexed by the Autism Quotient Communication subscale) showed a larger N400 to underinformative statements. In Experiment 2, this effect disappeared when the critical words were unfocused so that the local underinformativeness went unnoticed (e.g., “Some people have lungs that…”). Our results suggest that, while pragmatic scalar meaning can incrementally contribute to sentence comprehension, this contribution is dependent on contextual factors, whether these are derived from individual pragmatic abilities or the overall experimental context.

Keywords: Language comprehension, pragmatics, scalar quantifiers, underinformativeness, electrophysiology, N400 ERP, individual differences, pragmatic abilities, Autism Quotient Communication Subscale

INTRODUCTION

According one of the key principles of pragmatics, addressees by default presume that speakers communicate efficiently by uttering messages that are informative (Grice, 1975; Sperber & Wilson, 1986). This so-called conversational maxim of quantity is based on the idea that communication has evolved as a cooperative effort, and it often implicitly shapes our communicative interactions (e.g., Engelhardt, Bailey & Ferreira, 2006; see also Clark, 1996). Of course, that does not mean that everything that we say or write is genuinely informative. We easily adjust our expectations to who we are talking to (e.g., children, people who know more or less than we do), reflecting the fact that what is informative or relevant to one individual might be trivial or irrelevant to another. Moreover, there is abundant literature to suggest that individuals can vary greatly in their abilities to produce and comprehend pragmatic language, which could mean that some people are simply more focused on the logic of utterances than others (e.g., Baron-Cohen, 2008).

Although Grice’s account of pragmatic principles was not intended to serve as a psychological model of cognitive processing (see Bach, 2005; Bezuidenhout & Cutting, 2002), it may be that the addressee’s default presumptions have important ramifications for how language is processed online (e.g., Wilson & Sperber, 2004). One way in which Grice’s maxim of quantity may play out in online sentence processing is by influencing the addressee’s expectations of what kind of words will come next (e.g., Federmeier, 2007; Van Berkum, 2009). For example, following the sentence fragment “Some people have…”, the addressee might expect the upcoming word to denote something that not all people have (e.g., ‘pets’, ‘tattoos’), instead of something that all people possess (e.g., ‘lungs’, ‘bodies’). As a result, one can hypothesize that trivially true, underinformative statements (e.g., “Some people have lungs”) incur semantic processing costs because they deviate from the addressee’s expectations. In the two experiments reported below, we determined to what extent Grice’s maxim of informativeness contributes to the incremental build-up of sentence meaning. Specifically, we explored differences in individual’s reliance on this maxim for interpretation, and also investigated the role of general contextual factors on the processing of underinformative utterances. We addressed these issues by examining the impact of underinformative versus informative scalar sentences (e.g. “Some people have lungs/pets....”) on the N400 event-related potential (ERP), an electrophysiological index of semantic processing (Kutas & Hillyard, 1980, 1984).

Ever since Aristotle’s science of logic, quantifiers and logical operators have been important windows into human reasoning, and have maintained a crucial role in logic and linguistics because of their association with truth-value (e.g., see Gamut, 1991). The scalar quantifier ‘some’ has received much attention because it allows for two disparate readings: a pragmatic interpretation and a logical interpretation. The pragmatic interpretation approximates to ‘some but not all’ or ‘only some’. This interpretation constitutes a conversational inference, by which language comprehenders attribute an implicit meaning beyond the logical or literal meaning. This inference is termed a scalar inference or scalar implicature because it is thought that comprehenders base this pragmatic interpretation on the assumption that the communicator had a reason for not using a more informative or stronger term on the same quantity scale (some < many < all; see Horn, 1972). In other words, comprehenders assume that the communicator would have said ‘all’ if he/she thought ‘all’ was true, and assume that the communicator says ‘some’ because he/she thinks that stronger expressions like ‘many’ and ‘all’ are false.

The logical interpretation approximates to ‘at least some’ or ‘some and possibly all’. This interpretation makes sense when communicators use the expression ‘some’ when they lack all the relevant information (for example, “Some guests are coming to my party, but not everybody has RSVPed yet”, in which case it is possible that many or all invitees will come to the party), or when they are not referring to a specific subset (e.g., “Some people were crossing the street”).

Importantly, the pragmatic and logical interpretation may yield different truth values. For a simple, informative statement like “Some people have pets”, each interpretation yields an outcome that is true with respect to world knowledge; it is true that ‘some but not all people have pets’, consistent with the pragmatic interpretation, and it is also true that there exist people with pets, consistent with the logical interpretation. However, for an underinformative statement like “Some people have lungs”, whereas the logical interpretation yields a true outcome (because people with lungs do exist), the pragmatic interpretation yields a false outcome (because all people have lungs, not just some). The fact that ‘some’ may yield disparate truth-values can be used to examine how language comprehenders apply their pragmatic knowledge during sentence comprehension and establish sentence truth-value (for reviews see Noveck & Reboul, 2008; Noveck & Sperber, 2007; Sedivy, 2007).

Theoretical accounts of how people deal with scalar quantifiers predominantly differ in whether they assume that scalar inferences are generated by default or whether scalar inferences are context-dependent (see also Geurts, 2009; Horn, 2006; Recanati, 2003). In what has been dubbed the Levinsonian account, scalar inferences are generated automatically upon encountering ‘some’. The idea behind this is that, because the pragmatic meaning of scalars is so dominant in our language use, it has become ‘lexicalized’ (see Levinson, 2000; for related accounts see Chierchia, 2004; Gadzar, 1979) such that the intended message can be efficiently communicated. The pragmatic meaning, however, can be cancelled when the subsequent context requires so. For example, upon encountering the sentence “John wanted some of the cookies”, addressees automatically generate the pragmatic interpretation and interpret the sentences as meaning John wanted some, but not all, of the cookies. However, at a later point, upon encountering the sentence “In fact, he wanted all of them”, they revise their initial interpretation to be consistent with the logical interpretation. According to this account, it is this undoing of the scalar inference that is costly.

In contrast, proponents of Relevance Theory have posited that the generation of scalar inferences is chiefly a function of whether the inference is required to meet the addressee’s standard of relevance (e.g., Sperber & Wilson, 1986; Carston, 1998). The logical interpretation of ‘some’ (i.e., “some and possibly all”) could very well lead to a satisfying interpretation of the utterance, but the discourse context may require the addressee to derive a scalar inference to arrive at the pragmatic interpretation. Since this pragmatic interpretation involves ‘narrowing’ (negation of the stronger expressions ‘many’ and ‘all’), it constitutes a fully fledged inferential process which requires processing time and effort beyond the ‘easier’ logical interpretation.

Neither the Levinsonian framework nor Relevance Theory constitutes a psychological model of scalar inferences with explicit implications for processing. Yet, experimental psychologists have tried to infer testable predictions about the time course of scalar inferences. It has been argued that if scalar inferences are generated automatically, as advocated in the Levinsonian account, they are also generated relatively rapidly and their cancellation would incur additional processing costs (e.g., Bott & Noveck, 2004). In contrast, if scalar inferences are truly context-dependent, then they would incur processing costs in situations where they are not licensed by the context. According to Breheny, Katsos and Williams (2006), Relevance Theory predicts that in a neutral context (i.e., without a discourse context that biases towards either a logical or a pragmatic interpretation), no scalar inference will initially be computed, and only when the logical interpretation is deemed insufficient will addressees invest additional cognitive effort to generate a scalar inference.

To examine the time course for the generation of scalar inferences, behavioral research on scalar inferences has often used the sentence-verification paradigm (e.g., Bott & Noveck, 2004; Noveck, 2001; Noveck & Posada, 2003; Feeney, Scafton, Duckworth & Handley, 2004; Pijnacker, Hagoort, Buitelaar, Teunisse & Geurts, 2008; De Neys & Schaeken, 2007; for reviews, see Bezuidenhout & Morris, 2004; Noveck & Reboul, 2008; Noveck & Sperber, 2004, 2007; Huang & Snedeker, 2009; Sedivy, 2007). In sentence-verification tasks participants are asked to judge the truth of a statement, and in speeded sentence-verification tasks participants are asked to do this as fast as possible. Because the logical and pragmatic interpretation of informative sentences yield identical truth-values, the dependent measure of whether a scalar inference has been made is whether participants respond ‘false’ to an underinformative scalar statement (e.g., “Some people have lungs”). An often reported finding is that participants who respond ‘false’ to underinformative sentences are slower than those who respond ‘true’ (e.g., Bott & Noveck, 2004; Noveck & Posada, 2003; Rips, 1975). This is the case regardless of whether participants are explicitly instructed to respond ‘false’ or whether they spontaneously decide to do so (e.g., Bott & Noveck, 2004). These results have been interpreted as suggesting that scalar inferences are associated with additional processing costs and result from a delayed decision process (e.g., Bott & Noveck, 2004; Noveck & Posada, 2003; Noveck & Reboul, 2008).

Although using a sentence-verification task makes intuitive sense when dealing with truth-value, its interpretation is subject to a number of important caveats, as has already been noted by several researchers (Feeney et al., 2004; Grodner et al., 2010; Huang & Snedeker, 2009). For example, evaluating the logical meaning of an underinformative sentence may be inherently easier than evaluating its pragmatic meaning because one needs only one or two examples to verify the logical meaning (one or two people that have lungs) whereas one may need to do a more extended analysis to falsify the pragmatic meaning (e.g., search of, and failing to find counterexamples in memory; see also Grodner et al., 2010; Huang & Snedeker, 2009). Thus, it may not necessarily be the case that generating the pragmatic meaning requires additional processing effort and time, but rather refuting it. Another important concern is that speeded sentence verification is a relatively unnatural task that may encourage participants to ignore their pragmatic knowledge (Feeney et al., 2004), and it is hardly representative of how people process language in everyday life. Importantly, people who do generate scalar inferences are also slower in other conditions (e.g., Noveck & Posada, 2003), suggestive of a more general difference in task-related strategic processing. Finally, reaction times in verification tasks are generally quite slow, over 600 ms when statements are presented word by word (e.g., Noveck & Posada, 2003, Bott & Noveck, 2004) or even in the order of seconds when sentences are presented as a whole (e.g., Pijnacker et al., 2008; De Neys & Schaeken, 2007). In this regard, the results from verification tasks should be taken to reflect the combination of early stages of language processing as well as the output of downstream decision processes that follow them (e.g. Kounios & Holcomb, 1992).

Recently, researchers have overcome these problems by using a more indirect, high temporal resolution measure of scalar processing – the visual-world paradigm. Using this paradigm, Huang & Snedeker (2009) recorded eye-movements while participants received auditory instructions such as “Click on the girl that has some of the socks” or “Click on the girl that has all of the soccer balls” in the presence of a display in which one girl had two socks from the four socks that were present in the display, and another girl had all three soccer balls that were present in the display. The temporary referential ambiguity in the instruction at the point of ‘some’ could, in principle, be resolved immediately if participants made a scalar inference that would restrict ‘some’ to a proper subset. Participants, however, were substantially delayed, to ‘some’, but not when the instruction contained the word ‘all’. Based on this observation, Huang and Snedeker argued that ‘pragmatic’ scalar inferences are delayed relative to the ‘semantic’ logical interpretation (see also Bott & Noveck, 2004; Breheny et al., 2006; De Neys & Schaeken, 2007; Noveck & Posada, 2003).

However, Grodner and colleagues (Grodner, Klein, Carbary & Tanenhaus, 2010) note that ‘some’ is not unambiguously associated with a scalar inference (e.g., “Click on the girl with some socks” does not imply other socks are in the discourse), and that it was the partitive construction ‘of the’ that allowed for identification of the target in the Huang and Snedeker study. In contrast, for all, the quantifier itself was sufficient to identify the target. In a related study by Grodner et al. (2010) that circumvented these and some additional issues, scalar inference associated with pragmatic-some was not delayed relative to expressions that did not require a scalar inference. Thus, in contrast to the Huang & Snedeker (2009) results, the Grodner et al. results suggest that the pragmatic meaning of scalar expressions is rapidly available.

In the present study on scalar processing, we employed another indirect, high temporal resolution measure of language comprehension, namely Event-Related Potentials (ERPs). An important advantage of ERPs is that they provide both quantitative and qualitative information about language processing well in advance of (and without the principled need for) an explicit behavioral response (e.g., Van Berkum, 2004). In particular, we focus on the N400 ERP component (Kutas & Hillyard, 1980, 1984; see Kutas, Van Petten & Kluender, 2006, for review), a negative deflection in the ERP that emerges somewhere between 150 and 300 milliseconds after the onset of a word and that peaks at about 400 ms, with a maximum over the back of the head (i.e., electrodes at parietal locations). The N400 is, in principle, elicited by every content word, and its amplitude decreases in size and in a gradual manner when the word fits the context better (e.g., Kutas et al., 2006; Van Berkum, Brown, Zwitserlood, Kooijman, & Hagoort, 2005). A differential effect of two conditions on the N400 amplitude is referred to as an N400 effect. The functional significance of the N400 is still under debate (e.g., Kutas et al., 2006; Lau, Phillips & Poeppel, 2008; Van Berkum, 2009), but there is a general consensus that its amplitude reflects the fit between the lexical-semantic meaning of an incoming word and the interaction between linguistic context (at the level of single words, sentences and discourse) with information stored in memory (e.g., semantic memory, real-world knowledge and pragmatic knowledge of what a speaker is likely to say), henceforth referred to as ‘semantic fit’¹. The results from recent ERP studies have shown that the interaction between context and real-world knowledge can lead people to generate expectations about the semantic properties of specific upcoming words (e.g., De Long, Urbach, & Kutas, 2005; Federmeier, 2007; Van Berkum, 2009; Van Berkum et al., 2005), although it may be that, under other circumstances, the three-way mapping process is initiated only once the word is encountered. Importantly, in a recent study on negation processing, we showed that the N400 ERP is also sensitive to the informativeness of an utterance (Nieuwland & Kuperberg, 2008). In this study, participants read sentences that were true but underinformative due to pragmatically unlicensed negation (e.g., “Bulletproof vests aren’t very dangerous…”, in which case negation is used to deny something that makes no sense to begin with, namely that bulletproof vests are dangerous). Critical words (‘dangerous’) in these sentences elicited an increased N400 responses in the same way that false sentences did. In contrast, true sentences that contained pragmatically licensed negation (e.g., “With proper equipment, scuba-diving isn’t very dangerous…”) elicited N400 responses that were indistinguishable from those elicited by true affirmative sentences (e.g., “With proper equipment, scuba-diving is very safe…”). These results suggest that pragmatic knowledge of what is an informative thing to say influences an early stage of semantic processing, and may even contribute to building up broad pragmatic expectancies about what upcoming words are likely to be encountered.

There has been one previous study investigating whether the N400 is modulated by scalar inferences. Noveck and Posada (2003) recorded readers’ electrophysiological responses to sentence-final words in underinformative sentences (e.g., “Some elephants have trunks”), patently false sentences (e.g., “Some crows have radios”) and patently true sentences (e.g., “Some houses have bricks”). Similar to previous behavioral studies, participants were asked to make a speeded sentence verification response following each sentence. The results indicated that patently true and patently false sentences elicited a larger N400 ERP than underinformative sentences, and that the N400 responses to underinformative sentences were not modulated by whether participants responded true or false to these sentences. Consistent with previous behavioral findings, the reaction time data indicated that those participants who made scalar inferences (i.e. responded ‘false’ to underinformative sentences) were much slower to respond than those who followed a literal interpretation (i.e., responded ‘true’ to underinformative sentences). Critically, however, participants who made scalar inferences were much slower in all conditions, suggesting that these participants were using a more cautious strategy overall (see Feeney et al., 2004, for a related discussion). Noveck and Posada interpreted the smaller N400 for underinformative sentences, in combination with the slow time course of scalar implicatures, as being inconsistent with a Levinsonian account. They also suggested that scalar implicatures may likely be the product of a post-semantic decision process, that, once the critical word has been encountered, computes the truth-value of the complete proposition, whereas the initial stage of semantic processing after the critical word is determined only by simple lexical-semantic relationships (e.g., see also Fischler et al., 1983; Kounios & Holcomb, 1992). Later accounts by Noveck and colleagues suggest that, under certain conditions, the pragmatic scalar meaning may be generated without having to traverse through a logical interpretation first (Noveck & Sperber, 2007). However, the general idea that pragmatic processing costs are incurred after lexico-semantic processing is complete has persisted in some models of language processing (e.g. Bornkessel-Schlesewsky & Schlesewsky, 2008; Cutler & Clifton, 1999; Fodor, 1983; Forster, 1979; Regel, Gunter & Friederici, 2010).

Several problems with interpreting the initial ERP study by Noveck and Poseda. First, the materials in the different conditions were not matched or counterbalanced, and the words were presented in at a very fast pace (a presentation duration of 200 ms per word and an inter-word interval of 40 ms, which is about half of what is customarily used in ERP research using serial visual presentation²). Second, they employed a sentence-verification task that may have evoked decision-related positive ERPs that overlap in time and scalp distribution with the N400, and that may obscure modulations of the N400 (e.g., Kuperberg, 2007). In light of these concerns, it is important to note that patently false sentences did not evoke larger N400 responses than patently true sentences, whereas violations of real-world knowledge have consistently been associated with larger N400 responses in other studies (e.g., Fischler, Bloom, Childers, Roucos & Perry, 1983; Fischler, Childers, Achariyapaopan, & Perry 1984; Hagoort, Hald, Bastiaansen & Petersson, 2004; Hald, Steenbeek-Planting & Hagoort, 2007; Nieuwland & Kuperberg, 2008). This is problematic because these violations were included to establish a benchmark comparison for the main results.

In the current study, we addressed some of these concerns and used ERPs to examine how rapidly different individuals use their pragmatic knowledge of what is an informative versus uninformative thing to say during the processing of scalar sentences. We compared ERP responses elicited by critical words in underinformative scalar statements (e.g., ”Some people have lungs, …”) to those elicited by critical words in informative scalar statements (e.g., ”Some people have pets, …”, see Table 1 for more examples). If the pragmatic meaning of weak scalar quantifiers can be used incrementally during sentence comprehension (i.e., scalar inferences are made on-line), this may guide expectations about upcoming words so that readers and listeners will expect new input to be informative (e.g., Crain & Steedman, 1985; Altmann & Steedman 1988; Tanenhaus & Trueswell, 1995; see also MacDonald, Pearlmutter, & Seidenberg, 1994). Given that the N400 is sensitive to how well a word fits the context based on both semantic and pragmatic constraints (Coulson, 2004; Kutas et al., 2006; Nieuwland & Kuperberg, 2008; Van Berkum, 2009; Van Berkum, Van den Brink, Tesink, Kos, & Hagoort, 2008), this incremental account predicts that critical words in an underinformative statement would yield a larger N400 than in an informative statement.

TABLE 1.

Examples sentences from Experiment 1. Critical words are underlined for expository purpose only.

Underinformative/Informative

Some people have lungs/pets, which require good care.

Some rock bands have musicians/groupies, sometimes with drug problems.

Some gangs have members/initiations, and often strong hierarchy too.

Relatively poor/good semantic fit

Literature classes sometimes read papers/poems as a class.

Wine and spirits contain sugar/alcohol in different amounts.

Fillers

Many people catch the flu, especially in the winter.

Many vegetarians eat bean curd, which is rich in protein.

Open in a new tab

In contrast, if the pragmatic meaning of weak scalar quantifiers is not readily available when readers encounter the critical word, then the N400 ERP would not be sensitive to whether the statement is informative or underinformative. Rather, sentence processing and modulation of the N400 may be driven purely by lexico-semantic relationships (e.g., Otten & Van Berkum, 2007; Van Petten, 1993; Van Petten, Weckerly, McIsaac, & Kutas, 1997; for review, see Kutas et al., 2006). Because critical words in the underinformative condition (e.g., ‘lungs’) had a stronger lexical-semantic relationship to the main noun phrase in the preceding phrase (e.g., ‘people’) than in the informative condition (supported by their higher values on a Latent Semantic Analysis, LSA Landauer & Dumais, 1997), see Methods section), this would predict a smaller N400 to informative than non-informative sentences (as shown by Noveck and Poseda, 2003). This prediction also follows from Grice’s original account (for discussion see Geurts, 2009), and is generally consistent with models of language comprehension that assume that pragmatic factors come into play after an initial stage of ‘context-free’, linguistic-semantic processing (e.g. Fodor, 1983; Forster, 1979).

Previous studies have reported that individuals can vary significantly in whether and how they apply their pragmatic knowledge (e.g., Joliffe & Baron-Cohen, 1999; Musolino & Lidz, 2006; Noveck, 2001; Schindele, Lüdtke & Kaup, in press; Stanovich & West, 2001; Tager-Flusberg, 1981). Moreover, there have been several reports of individual differences in scalar inference generation (e.g., Bott & Noveck, 2004; Feeney et al., 2004; Noveck & Posada, 2003), suggesting that different people may preferentially and consistently adopt either a literal or a pragmatic interpretation when asked to evaluate underinformative sentences. Our hypothesis, which we will describe in more detail below, is that individuals with good real-world pragmatic skills are, at least initially, relatively more sensitive to the pragmatic ‘violation’ of underinformativeness and therefore more likely to show a pragmatic N400 effect, whereas processing in people with poorer real-world pragmatic skills is more likely to be driven by pure lexico-semantic association.

As a caveat, inferences regarding the full extent of incremental scalar processing based on our paradigm are limited. As opposed to studies that have used the visual-world paradigm, our study was not designed to examine whether scalar inferences are generated immediately upon encountering the scalar quantifiers. A modulation of the N400 by informativeness in our study could be taken either as evidence that the processing consequences of the scalar quantifier either are rapidly computed upon encountering the critical word, or were perhaps computed before encountering the critical word. As argued by Van Berkum (2009), there are good reasons to assume that a pragmatic modulation of the N400 does not directly reflect a fully compositional enrichment process, but more likely indicates that the semantic and pragmatic consequences of the preceding discourse have been computed to serve as an interpretive background to retrieve word meaning (see also Kuperberg, Paczynski & Ditman, in press). Although our study was not specifically designed to examine ERP responses to scalar quantifiers, we will report exploratory analyses that address these issues.

EXPERIMENT 1

In the first of our two experiments, we examined electrophysiological responses to critical words in underinformative statements versus informative scalar statements, and used this measure to investigate individual differences in pragmatic processing. If scalar pragmatic inferences are generated incrementally during online sentence processing, critical words that render a statement trivial or underinformative should lead to additional semantic processing costs, and should elicit a larger N400 than critical words in informative statements – a pragmatic N400 effect (see also Nieuwland & Kuperberg, 2008). If, on the other hand, pragmatic scalar information is not used incrementally during online processing, the N400 should not be larger to critical words in underinformative statements. In fact, given the closer lexico-semantic associations in underinformative than in informative sentences (people-lungs vs. people-pets), the N400 may even be relatively attenuated in underinformative sentences.

We also hypothesized that there may be individual variation in these patterns of N400 modulation, that may be predicted by variation in participants’ abilities to produce and comprehend pragmatic aspects of language in the real world (e.g., Baron-Cohen, Tager-Flusberg & Cohen, 2000; Happé, 1993; Schindele et al., 2008; Tager-Flusberg, 1981, 1985). We therefore obtained an independent measure of pragmatic language abilities of our participants in everyday life through the Communication subscale of the Autism-Spectrum Quotient questionnaire (the AQ; Baron-Cohen, Wheelwright, Skinner, Martin & Clubley, 2001) that quantifies an individual’s pragmatic skills on a continuum from autism to typicality. Of the five AQ subscales, the Communication subscale taps into pragmatic abilities most directly. Some examples of items from this subscale are “Other people frequently tell me that what I’ve said is impolite, even though I think it is polite”, “I find it hard to ‘read between the lines’ when someone is talking to me”, and “I am often the last to understand the point of a joke”.

We predicted that individuals with good pragmatic abilities (as indexed by a low score on the AQ Communication subscale), would be relatively more sensitive to the pragmatic ‘violation’ of underinformativeness and more likely to show a pragmatic N400 effect, as compared to less pragmatically skilled individuals (see Schindele et al., 2008, Pijnacker et al., 2008, for related hypotheses in participants with high-functioning autism or Asperger’s syndrome). This sensitivity may play out in several different ways. For example, individuals with good pragmatic abilities might generate pragmatic inferences more consistently, generate more robust inferences, they might be better at evaluating incoming words for informativeness, or perhaps even have a different task set than people with poor pragmatic skills. In the current study we cannot distinguish between these or other possibilities. Nevertheless, modulation of a pragmatic N400 effect by pragmatic abilities could provide evidence that such everyday communication problems may be, in part, driven by an impaired incremental use of pragmatic knowledge during language processing.

In order to examine the specificity of these potential individual differences, we also included sentences that did not contain scalars, but that contained a word that had a relatively good semantic fit versus relatively poor semantic fit to the preceding sentence context based on real-world knowledge; see Table 1 for examples). We predicted that words that were incongruous with real-world knowledge³ would produce a robust N400 effect compared to words that were congruous with real-world knowledge (e.g., Kutas & Hillyard, 1984) in all individuals, regardless of their AQ-Communication scores. This allowed us to dissociate individual differences in incrementally recruiting pragmatic knowledge from the more general recruitment of real-world knowledge during online processing.

METHODS

Participants

Thirty-one right-handed Tufts students (17 males; mean age = 20.2 years) gave written informed consent. All were native English speakers, without neurological or psychiatric disorders.

Materials

We constructed 70 sentence pairs such that the underinformative and informative versions of each sentence pair were identical except for the critical word. Each sentence consisted of two clauses, and the first clause (the quantifier clause) always started with the quantifier ‘some’ and always ended with a comma after the critical word. We selected critical words so that replacing ‘some’ by the quantifier ‘all’ would yield a true statement in the underinformative condition (e.g., “All people have lungs”), but a false statement in the informative condition (e.g., “All people have pets”). The second clause always contained at least three words and provided additional information about the critical word, the main NP in the scalar clause (e.g., ‘people’) or the scalar clause as a whole, and was created so that the complete sentence constituted a logically true statement in each condition. Critical words in the two conditions were approximately matched for average length in number of letters (underinformative, informative, M = 6.7/7.0, SD = 1.8/2.0) and log frequency (Francis & Kucera, 1976; underinformative, informative, M = 1.73/1.91, SD = 2.29/2.03). Semantic similarity values were calculated for the critical words within the underinformative and informative sentences using Latent Semantic Analysis (Landauer & Dumais, 1997; Landauer, Foltz, & Laham, 1998; available on the Internet at http://lsa.colorado.edu). As expected, underinformative words yielded a higher LSA value than informative words (underinformative, informative; M =.33/.17, SD =.23/.18; t(138) = 4.58, p < .001). As noted in the Introduction, higher LSA values are generally associated with smaller N400 amplitudes compared to lower LSA values, because the LSA values reflect in part the amount of lexico-semantic priming a word receives from the preceding context.

For the semantic fit manipulation, we constructed another 70 sentence pairs that were identical except for the critical word. Critical words were selected that were relatively congruous or incongruous to the sentence with regard to world knowledge (see Table 1 for examples). Critical words in the two conditions were matched for average length in letters (congruous, incongruous, M = 6.4/6.3, SD = 2.1/1.7) and log frequency (Francis & Kucera, 1982; congruous, incongruous, M = 1.46/1.50, SD = 1.74/1.88). Semantic similarity values were calculated for the congruous, incongruous words using Latent Semantic Analysis. Good semantic fit words yielded a higher LSA value than poor semantic fit words (congruous, incongruous, M =.22/.14, SD =.11/.08; t(138) = 5.31, p < .001). At least two words followed the critical words before the sentence ended.

We also created 35 filler sentences that each had a similar sentence structure as the scalar sentences but that always started with the quantifier ‘many’, and involved a simple and true statement (e.g., “Many vegetarians eat bean curd, which is rich in protein.”).

We created two counterbalanced lists so that each sentence appeared in only one condition per list, but in all conditions equally often across lists. Within each list, items were pseudorandomly mixed with the 70 sentences containing a semantic fit manipulation (35 containing a relatively good fitting critical word, 35 containing a relatively poor fitting critical word) and the 35 filler sentences to limit the succession of identical sentence types, while matching trial-types on average list position.

The Autism-Spectrum Quotient

The AQ (Baron-Cohen et al., 2001) is a self-administered questionnaire that is designed to measure the extent to which adults with normal intelligence possess traits associated with Autism Spectrum Disorder (ASD). Although this scale is not a diagnostic measure, its discriminative validity as a screening tool has been clinically tested (Woodbury-Smith, Robinson, Wheelwright, & Baron-Cohen, 2005). The test consists of 50 items, made up of 10 questions assessing five subscales: Social Skill (e.g., “I would rather go to a library than a party”), Communication (e.g., “I frequently find that I don’t know how to keep a conversation going”), Imagination (e.g., “When I’m reading a story, I find it difficult to work out the characters’ intentions”), Attention To Detail (e.g., “I usually notice car number plates or similar strings of information”), and Attention-Switching (e.g., “I frequently get so absorbed in one thing that I lose sight of other things”). Half the questions are worded to elicit an ‘agree’ response and the other half a ‘disagree’ response, addressing demonstrated areas of cognitive characteristics in ASD (DSM-IV, 1994; Baron-Cohen et al., 2001). Higher scores on the AQ indicate stronger presence of traits associated with ASD. A score of 32+ appears to be a useful cutoff for distinguishing individuals who have clinically significant levels of autistic traits (Baron-Cohen et al., 2001; the maximum score of the participants in our study was 30). Such a high score on the AQ however does not mean that an individual has autism, because a diagnosis is only merited, based on diagnostic measures such as the DSM-IV (1994), ADI-R (Lord, Rutter & Couteur, 1994) or ADOS-G (Lord et al., 2000), if the individual is suffering a clinical level of distress as a result of their autistic traits. In the current study, the AQ was administered in a quiet room subsequently to the ERP experiment, and took each participant about 10 minutes.

Procedure

Participants silently read sentences, presented word-by-word and centered on a computer monitor, while minimizing eye-movements and blinks. There was no task other than reading for comprehension. To parallel natural reading times (Legge, Ahn, Klitz & Luebker, 1997), all words were presented using a variable presentation procedure (Otten & Van Berkum, 2008; see also Nieuwland & Kuperberg, 2008). Word duration in ms was computed as ((number of letters × 27) + 187), with a 10 letter maximum. Also, to mimic natural reading times at clause boundaries (e.g., Hirotani, Frazier & Rayner, 2006; Legge et al., 1997; Rayner, Kambe & Duffy, 2000), critical words (which were followed by a comma) were presented for an additional 227 ms, and sentence-final words for an additional 500 ms. All inter-word-intervals were 121 ms. Following sentence-final words, a blank screen was presented for 500 ms, followed by a fixation mark at which subjects could blink and self-pace on to the next sentence by a right-hand button press. Participants were given six short breaks. Total time-on-task was approximately 40 minutes. After the ERP experiment, each subject was allowed a short break to wash up and was then administered a brief exit-interview, followed by the Autism-Spectrum Quotient questionnaire.

In the exit-interview, participants received a booklet that contained 6 pages and were instructed to answer the question from the booklet page-by-page without looking at the subsequent pages. On page 1, subjects were asked to report whether they noticed anything about the sentences they read and what research question(s) they thought the experiment was about. On page 2, an example of an informative scalar sentence was given, and participants reported whether they thought that sentences starting with ‘Some’ stood out, what they thought the purpose of these sentences was, and what research question these sentences involved. On page 3, subjects reported whether they thought that some of the sentences in the experiment sounded odd and provided a brief explanation why they thought this. On pages 4 and 5, subjects were presented with 10 different scalar statements, including informative and underinformative scalar sentences truncated after the CW as well as longer sentences that contained locally informative or under informative phrases. Subjects were asked to rate whether each sentence was true (1=false, 5=true) and how normal they would find it if somebody said this (1=odd, 5=normal). On page 6, subjects were informed that a sentences like “Some people have lungs” could be rated as false (because the sentence implies that most people do not have lungs) or true (because there are at least some people in the world that do have lungs). The subjects were asked to report whether they thought during the experiment about whether these sentences were true or false, whether they during the experiment ‘treated’ these sentences as true or false, and how consistently they did this (1=very inconsistently, 5=very consistently).

EEG Recording

The electroencephalogram (EEG) was recorded from 29 tin electrodes held in place on the scalp by an elastic cap (Electro-Cap International, Inc., Eaton, OH, USA). Electrode locations included Fz, Cz, Pz, Oz, Fp1/2, F3/4, F7/8, FC1/2, FC5/6, C3/4, T3/4, T5/6, CP1/2, CP5/6, P3/4, P7/8, O1/2, and 2 additional EOG electrodes; all were referenced to the left mastoid). The EEG recordings were amplified (band-pass filtered at 0.01 Hz–40 Hz) and digitized at 200 Hz. Impedance was kept below 5 kOhm for EEG electrodes. Prior to off-line averaging, single-trial waveforms were automatically screened for amplifier blocking and muscle/blink/eye-movement artifacts over 850 ms epochs (starting 100 ms before CW onset). Two participants were excluded due to excessive artifacts (mean trial loss > 50%). For the remaining 29 participants, average ERPs (normalized by subtraction to a 100 ms pre-stimulus baseline) were computed over artifact-free trials for CWs in all conditions (mean trial loss across conditions 11%, range 0–42%, without substantial differences in mean trial loss across conditions).

Statistical analysis

For all analyses reported below, the Greenhouse/Geisser correction was applied to F tests with more than one degree of freedom in the numerator. Note that due to the large number of trials needed for averaging in ERPs (which reduces the probability that the results hinge on just a few odd items), statistics are only reported for by subjects analyses, and analyses by items are not included.

RESULTS

Main effect of informativeness

Critical words elicited very similar N400 responses in the underinformative and the informative statements (see Figure 1, left panel). Because modulation of the N400 ERP is generally maximal at posterior electrodes (e.g., Kutas et al., 2006), we divided all electrodes into anterior electrodes (F3/4, F7/8, F9/10, FC1/2, FC5/6, FP1/2, FPz, Fz) and posterior electrodes (Pz, Oz, CP1/2, CP5/6, P3/4, P7/8, O1/2) for subsequent analyses. Using mean amplitude in the 350 to 450 ms time window, a 2 (informativeness: informative, underinformative) × 2 (AP distribution: anterior, posterior) repeated measures analysis of variance (ANOVA) revealed that there was no statistically significant difference between the ERP responses to informative and underinformative statements, and no interaction effect between informativeness and AP distribution.

*Left panel*: Grand-average event-related potential (ERP) waveforms elicited by critical words in underinformative (dotted lines) and informative (solid lines) statements from Experiment 1, shown at electrode locations Cz, Pz, and Oz. In this and all following figures, negativity is plotted upwards. *Middle panel*: Grand average ERPs elicited by critical words in underinformative and informative statements per AQ Communication group in Experiment 1, and corresponding scalp distributions of the mean difference effect (underinformative minus informative sentences) in the 350- to 450 ms analysis window. *Right panel*: Correlation between N400 effect and AQ Communication score.

AQ-Comm score and ERP responses to informativeness

AQ scores ranged from 9 to 30 (M=21, SD=7.04). To explore the role of pragmatic abilities, we first grouped the participants into low AQ-Comm (N=15) and high AQ-Comm (N=14) groups based on the median split of scores on the Communication subscale. AQ-Comm score for the low AQ-Comm group ranged from 0 to 5 (M=2.33, SD=.51; 7 males and 8 females, mean age 20.9 years, mean total AQ score 15.8), and from 6 to 9 for the high AQ-Comm group (M=7.2, SD=.28; 8 males and 6 females, mean age 19.3 years, mean total AQ score 26.5). The two AQ-Comm groups showed statistically significant differences in AQ-Comm score (t(27)=8.34, p<.001) and in total AQ score (t(27)=6.24, p<.001), as well as in age (t(27)=2.49, p<.05; when entered into the subsequent analyses as a covariate, the factor age, however, did not change the patterns of results.

Grand average ERPs for the two groups are displayed in Figure 1 (middle panel). Using mean amplitude in the 350 to 450 ms time window, the overall ANOVA revealed a significant 2 (informativeness: informative, underinformative) × 2 (AQ-Comm Group: low AQ-Comm, high AQ-Comm) interaction effect when using all electrodes (F(1,27)=9.45, p=.005). There was no significant 3-way interaction with AP distribution (F(1,27)=2.19, p=.15), but the Informativeness by AQ-Comm group interaction effect was statistically significant when using only posterior electrodes (F(1,27)=11.54, p=.002), but only marginally significant when using anterior electrodes (F(1,27)=3.3, p=.07). This predominantly posterior distribution of N400 modulation is consistent with the N400 literature (e.g., Kutas et al., 2006).

Simple main-effect analysis for the groups separately, using posterior electrodes only, showed that underinformative statements elicited larger N400 responses than informative statements in the low AQ-Comm group (F(1,14)=5.57, p=.033, CI −.82 ± .75), whereas informative statements elicited larger N400 responses than underinformative statements in the high AQ-Comm group (F(1,13)=6.12, p=.028, CI −1.38 ± 1.2). There was no statistically significant effect of informativeness in the two AQ-Comm groups separately when taking into account anterior electrodes only (Fs<1, n.s.).

As can be seen from Figure 1, there appeared to be differential effects of informativeness for the two groups before the 350–450 ms time window. We therefore performed additional 2 (informativeness: informative, underinformative) × 2 (AQ-Comm Group: low AQ-Comm, high AQ-Comm) ANOVAs for the 50–150, 150–250 and 250–350 time windows. These revealed some significant effects within early time windows (50–150 ms in the low AQ-Comm group, 150–250 ms in the high AQ-Comm group; see Appendix A for full report, which can be found at http://www.nmr.mgh.harvard.edu/kuperberglab/materials.htm). We were concerned that these early effects of informativeness reflected an artefactual side effect of dividing subjects on the basis of their AQ-Comm score. It is well-known that with limited numbers of EEG trials going into the average of a single subject, single-subject ERPs constitute unknown mixtures of critical ERP effects and residual EEG background noise which could, in principle, explain the early onset ERP differences. We therefore repeated analyses using a longer, 500 ms pre-CW baseline thus reducing noise in the baseline time window (and consequently, in the post-baseline ERP signal). ERP difference effects that truly are the result of the experimental manipulation should survive this longer baseline analysis. The corresponding figures for these analyses can be found at the website as referenced above. After rebaselining, the early effects in the 50–150 and 150–250 ms windows disappeared but left the main pattern of results in the 250–350 and 350–450 ms windows unchanged (see Appendix A). Additional analyses for the post-450 ms time windows using the original baseline as well as the new baseline can also be found on our website.

Correlation analysis for AQ-Comm scores and ERP responses to informativeness

We also performed a correlation analysis that took into account the full range in individual AQ-Comm scores, and revealed a negative correlation between AQ-Comm score and the mean ERP difference score calculated as underinformative minus informative in the 350–450 ms time window at posterior electrodes (Pearson’s r = −.53, p=.003; see Figure 1, right panel). This correlation effect was also present for total AQ score (r = −.55, p=.002), the Social Skill subscale score (r = −.45, p=.014) and Attention-Switching subscale score (r = −.55, p=.002), but was not significant for scores on the subscales Imagination (r = −.21, p=.29) and Attention To Detail (r =.17, p=.39). We should note that the Attention-Switching subscale and the Communication subscale were also the strongest interrelated subscales, so the effects of these subscales are hard to tease apart.

ERP responses to informativeness and the role of LSA

As mentioned in the Introduction, the content words in underinformative statements co-occur in language relatively more frequently than those in the informative statements, as reflected by their differences in LSA values. However, not each underinformative statement from each sentence pair had a larger LSA value than its informative counterpart. This allowed us to separate our items into one set that had a relatively small LSA difference between informative and underinformative sentences (LSA(underinformative-informative), M = −0.02, SD = 0.12), and one set that had a relatively large LSA difference across conditions (M =0.34, SD =0.18). By computing ERPs separately for these two sets for each group, we investigated the effect of informativeness while controlling for lexical-semantic factors.

The corresponding figures for these analyses can be found at (http://www.nmr.mgh.harvard.edu/kuperberglab/materials.htm). These plots reveal clear differences between the low and high AQ-Comm groups in N400 modulation by LSA and informativeness. Analyses focusing on N400 peak amplitude modulations across posterior electrodes in the 350–450 ms time window showed that the informativeness by LSA difference interaction effect was significant in the high AQ-Comm group (F(1,13)=5.38, p=.037), but not in the low AQ-Comm group (F(1,14)=.02, p=.90). Follow-ups showed that, in the low AQ-Comm group, critical words in underinformative statements elicited a larger N400 than those in informative statements, both when there was a relatively small and a relatively large LSA difference between conditions (small difference, F(1,14)=2.37, p=.043, CI −.84 ± .76; large difference, F(1,14)=2.19, p=.052, CI −.80 ± .77). In the high AQ-Comm group, however, underinformative statements elicited a lower N400 than informative statements only when there was a relatively large LSA difference (F(1,13) =4.01, p= 0.001, CI −2.25 ± 1.21), but not when there was a relatively small LSA difference (F(1,13) =.21, p= 0.834, CI −.17 ± 1.76).

In sum, whereas we found a typical modulation of LSA in the high AQ-Comm group, the pragmatic N400 effect in the low AQ-Comm group was insensitive to LSA.

Group differences in ERP responses to sentence-final words

We also examined the ERP responses to sentence-final words in underinformative and informative statements between the two AQ-Comm groups (see Figure 2). Statistical analyses were carried out using mean amplitude in the 300 to 500 ms. The sentence-final words involved different word categories, and there may have been differences in naturalness of the second clauses following informative versus underinformative statement. Our main interest in this comparison was therefore not the main effects of informativeness (positive ERPs to sentence-final words of underinformative than informative sentences across both groups, F(1,27)=20.28, p<.001, CI .96 ± .46), but rather the differences between the two AQ-Comm groups to the same set of stimuli. As shown in Figure 2, there was a clear differential ERP effect on the sentence-final words in underinformative and informative statements in the low AQ-Comm group, but less so in the high AQ-Comm group. This differential ERP effect appeared to have a slightly frontal distribution (i.e., inconsistent with an N400 effect scalp distribution), and may reflect additional sentence wrap-up processing. Across all electrodes, the overall ANOVA revealed a marginally significant informativeness by AQ-Comm group interaction effect (F(1,27)=3.88, p=.059) and follow-ups showed that the modulation by informativeness was significant in the low AQ-Comm group (F(1,27)=17.56, p=.001, CI 1.36 ± .70), but only marginally significant in the high AQ-Comm group (F(1,27)=3.64, p=.079, CI .54 ± . 60). A 2 (informativeness: informative, underinformative) × 2 (AP distribution: anterior, posterior) ANOVA revealed no interaction effect of informativeness with anterior-posterior distribution (F<1), and there was no significant interaction between informativeness, AQ-Comm group and distribution (F<1). Because the effect was prolonged, we repeated the above analyses in the 500–700 ms window and this yielded the same pattern of results.

Grand average ERPs elicited by sentence-final words in underinformative (dotted lines) and informative (solid lines) statements per AQ Communication group in Experiment 1, shown at electrode locations FPz, Cz, and Oz, and corresponding scalp distributions of the mean difference effect (underinformative minus informative sentences) in the 300- to 500 ms and the 500- to 700 ms analysis window.

Group differences in ERP responses to real-world congruous versus incongruous sentences

To determine the specificity of the group differences in ERP responses to underinformativeness, we also examined whether the groups differed in their N400 modulation to words that were congruous versus incongruous with real-world knowledge. We compared the modulation of the N400 by words with a relatively poor versus good fit based on real-world knowledge across the two groups. As can be seen from Figure 3, the modulation of the N400 was quite similar across the two groups. Using mean amplitude at posterior electrodes in the 350 to 450 ms time window, the overall 2 (Real world congruity: congruous, incongruous) × 2 (AQ-Comm Group: low AQ-Comm, high AQ-Comm) ANOVA revealed that the incongruous words evoked a larger amplitude N400 than congruous words (F(1,27)=19.28, p<.001, CI −1.35 ± .64) However, no Real world congruity by AQ-Comm Group interaction was observed (F(1,27)=1.77, p=.19). There was also no significant Real world congruity by AQ-Comm Group interaction in the adjoining 250–350 and 450–550 time windows (all Fs < 2, ns.). Consistent with the absence of this interaction, there was also no significant correlation between the N400 difference effect in the 350–450 ms time window and AQ-Comm score (Pearson’s r = −.29, p=.13).

*Left panel*: Grand average ERPs elicited by words that had a relatively poor (dotted lines) and relatively good (solid lines) semantic fit per AQ Communication group in Experiment 1, and corresponding scalp distributions. *Right panel*: Correlation between N400 effect and AQ Communication score.

Exploratory analyses of ERP responses to the scalar quantifiers

Although our experiment was not specifically designed to examine ERP responses to the scalar quantifiers, we performed an exploratory analysis to investigate whether there were differences between the two AQ-Comm groups in ERP responses to the sentence-initial scalar quantifiers ‘Some’ (the sentence-initial word of the experimental sentences) and ‘Many’ (the sentence-initial word in 35 filler sentences). The reasoning behind this analysis was that if the quantifiers themselves evoke differential pragmatic processing, then the differences in pragmatic abilities between the groups may already become apparent at the quantifier. We note that the quantifier ‘many’ can elicit a “not all” implicature as can ‘some’, so this comparison is not optimal for examining differences in pragmatic processing. However, because these quantifiers can be arranged on a scale of informativeness where ‘many’ is stronger than ‘some’, the ‘some’ implicature would include “not many” as well as “not all”. In this sense, and particularly in an experimental context in which both are repeatedly presented, one could argue that these scalar quantifiers are associated with implicatures that are of different strength.

The figures corresponding to this analysis can be accessed at http://www.nmr.mgh.harvard.edu/kuperberglab/materials.htm. In the high AQ-Comm group, ‘Many’, relative to ‘Some’ appeared to evoke a slightly more negative right-lateralized waveform at about 300–350 ms and a more positive frontally-distributed waveform at about 650–700 ms. There appeared to be no such effect in the low AQ group. We performed a series of repeated measures ANOVAs to test for the 2 (quantifier: some, many) by 2 (AQ-Comm group: low AQ-Comm, high AQ-Comm) interaction, in adjoining 50 ms time windows between 100 and 800 ms after quantifier onset, using all electrodes or only anterior or posterior electrodes. The only (marginally) significant interaction effect was found in the 650–700 ms window using anterior electrodes (F(1,27)=3.77, p=.063). Follow-up analyses confirmed that ‘many’ elicited more positive ERPs than ‘some’ in the high AQ-Comm group (F(1,13)=14.70, p=.002, CI −2.04 ± 1.15), but there was no difference between the two quantifiers in the low AQ-Comm group (F(1,14)=.07, p=.80, CI −.21 ± 1.68). In addition, this frontal positivity effect showed a marginally significant correlation with AQ-Comm score (Pearson’s r = .34, p=.073). There was also a marginally significant correlation between the frontal positivity effect and the differential ERP effect at the critical words, suggesting that participants who showed a larger frontal positive effect were less likely to show a pragmatic N400 effect later in the sentence (r = −.35, p=.06). The frontal positivity, however, did not predict the N400 modulation by real-world congruity (Pearson’s r = −.15, p=.46).

Exit interview

We examined whether the AQ-Comm groups differed in their exit-interview ratings for truth-value and naturalness. A 2 (AQ-Comm group: low AQ-Comm, high AQ-Comm) by 2 (informativeness: informative, underinformative) ANOVA revealed no group differences in the truth-value ratings and the naturalness ratings (all Fs<2). In addition, underinformative and informative statements received similar truth-value ratings (t<1) but different naturalness ratings (t(1,28)= 15.98, p < .001).

DISCUSSION

Across all participants, underinformative statements elicited N400 responses that were similar to those elicited by informative statements. However, there was marked heterogeneity across individuals in N400 modulation, with some individuals showing a larger N400 to critical words in underinformative than in informative statements, and others showing the opposite pattern of modulation (i.e., a larger N400 to critical words in informative than underinformative statements). Most importantly, these individual differences could be explained by taking into account individual variability in real-world pragmatic language ability. Individuals with few pragmatic language difficulties (as indexed by a low score on the AQ Communication subscale) were more sensitive to the pragmatic ‘violation’ of underinformativeness. This opposite pattern of activity was clear both in a median split analysis that dichotomized the two groups and in a correlation analysis that took into account the full range in individual AQ-Comm scores. Importantly, this N400 modulation by AQ-Comm score did not extend to the N400 responses to words with a relatively poor fit with respect to world knowledge, suggesting that AQ-Comm score was fairly specific in explaining the pattern of N400 modulation to the pragmatic violations. In addition, the two groups were differentially sensitive to lexical-semantic co-occurrence: whereas the low AQ-Comm group showed a pragmatic N400 effect independently of whether the underinformative and informative sentences were matched for LSA, the high AQ-Comm group’s ERP responses were modulated by LSA. Finally, we also explored ERP responses to the scalar quantifier ‘some’ versus’ many’. Although these quantifiers could be argued to evoke related (although not identical) pragmatic processes, rendering this comparison suboptimal for examining potential differences in pragmatic processing, we did find some preliminary evidence that pragmatic abilities influenced processing at the scalars themselves.

If one considers only the pragmatically skilled participants, our results show that pragmatically underinformative statements are associated with early semantic processing costs (see also Nieuwland & Kuperberg, 2008). This result suggests that the pragmatic meaning of a scalar quantifier can, in principle, be rapidly and incrementally incorporated during sentence comprehension, a finding that is consistent with models of language processing that incorporate an incremental contribution of pragmatic factors (Crain & Steedman, 1985; Altmann & Steedman, 1988; Tanenhaus & Trueswell, 1995) and with the results of studies from the visual world paradigm (Grodner et al., 2010).

In contrast to the more pragmatically skilled participants, however, the less pragmatically skilled participants showed no pragmatic N400 effect. Their processing was rather driven primarily by the relatively closer lexical-semantic relationships between individual words in these statements which overrode pragmatic factors. One possible interpretation of these results is that these individuals, who report difficulties with pragmatic abilities in everyday life, were simply incapable of generating scalar inferences. One could argue that this conclusion is in line with the notion from Relevance Theory that scalar inferences are not obligatory (see also Bott & Noveck, 2004; Noveck & Posada, 2003) but depend on constraints from the context and possibly from neuropsychological factors (see also Happé, 1993).

However, if one takes into account the ERP patterns elicited by sentence-initial scalars, a more complicated picture emerges. The exploratory analyses of ERP responses elicited by the sentence-initial scalar quantifiers suggest that pragmatic abilities influenced scalar statement processing already at the scalar quantifier. Perhaps counterintuitively, differential processing of the two different scalar quantifiers was most pronounced in the pragmatically less skilled participants. We will provide more in-depth discussion of these issues in the general discussion, but what these results suggest is that pragmatically less skilled participants may have been able to temporarily ignore or inhibit their pragmatic knowledge during the processing of the critical words (see Feeney et al., 2004; Handley & Feeney, in press), instead of being insensitive to pragmatic constraints (e.g., Schindele et al., 2008).

In sum, our results suggest that pragmatic constraints can have rapid effects during on-line sentence comprehension. When pragmatic constraints are taken into account, as in low AQ-Comm people, they may guide expectations about upcoming words through the pragmatic presumption of informativeness. But when these constraints cannot be used or they are ignored, as in the high AQ-Comm group, the effects of other constraints may surface, such as the effect of lexical-semantic relationships. In our second experiment, we examined the incremental processing of weak scalar quantifiers further by modulating the effect of pragmatic constraints through linguistic focus.

EXPERIMENT 2

INTRODUCTION

Whereas blatantly underinformative statements that violate pragmatic principles are relatively uncommon in everyday language (perhaps with the notable exception of utterances where underinformativeness is used as a humoristic device), temporarily underinformative statements are quite common. For example, whereas a phrase such as “Some people have eyes,” is unlikely to appear, a sentence such as “Some people have eyes that are different colors” is much more natural.

In the sentence “Some people have eyes,” the comma signals clausal wrap-up and the end of the quantifier scope. This puts the clause-final words ‘eyes’ clearly into focus (e.g., Birch & Rayner, 1997; Hirotani, Frazier & Rayner, 2006). In contrast, in “Some people have eyes that are different colors”, the scope of the quantifier encompasses the whole relative clause construction (‘eyes that are different colors’) and the focus of the utterance – the part of the statement that the communicator wants to emphasize and is most relevant to the addressee for evaluating sentence meaning – is not ‘eyes’ but ‘different colors’.

Research on the role of focus in language comprehension suggests that the processing of unfocused materials is dominated by ‘low-level’ lexical-semantic relationships rather than by ‘full-fledged’ compositional processing that is needed to establish sentence truth-value or real-world plausibility (e.g., Ferreira, Bailey & Ferraro, 2002; Sanford & Sturt, 2002). This is because readers and listeners generally devote less attention and processing time to unfocused material than to focused material (e.g., Cutler, Dahan & Van Donselaar, 1997; Frazier, Carlson & Clifton, 2006), so that unfocused materials receive an incomplete semantic and pragmatic analysis (so-called shallow processing; e.g., Sanford & Garrod, 1998).

In Experiment 2, a second set of participants read sentences like “Some people have eyes that are different colors”. The first clauses of these sentences were identical to those used in Experiment 1, but the comma was excluded and the clause was always followed by a relative clause (see Table 2, for examples). Thus the sentences were informative overall but the first clause could be considered ‘locally’ underinformative. Given the absence of the comma and the fact that all scalar sentences in Experiment 2 had this same structure, we expected the critical words to be out of focus and we hypothesized that they would therefore be processed more shallowly (e.g., Sanford & Sturt, 2002), and the ERP response would be dominated by simple lexical-semantic relationships rather than by the pragmatic presumption of informativeness. In other words, we predicted that locally underinformative statements would fail to evoke a pragmatic N400 effect. Rather, we predicted that the N400 would be reduced, relative to the informative statements, because of their closer lexical-semantic relationships. In addition, we predicted that this effect would not be modulated by the real-world pragmatic abilities (AQ-Comm score) of the participants because the relevant pragmatic constraints were now the same in the informative and underinformative statements.

TABLE 2.

Examples sentences from Experiment 2. Critical words are underlined for expository purpose only.

Underinformative/Informative

Some people have lungs/pets that are diseased by viruses.

Some rock bands have musicians/groupies with real drug problems.

Some gangs have members/initiations that are really violent.

Relatively poor/good semantic fit

Literature classes sometimes read papers/poems as a class.

Wine and spirits contain sugar/alcohol in different amounts.

Fillers

Many people catch the flu in the winter.

Many vegetarians eat bean curd as a source of protein.

Open in a new tab