Portions and sorts in Icelandic: an ERP study.

Brain & Language 136 (2014) 44–57

Contents lists available at ScienceDirect

Brain & Language journal homepage: www.elsevier.com/locate/b&l

Portions and sorts in Icelandic: An ERP study Matthew Whelpton a, Drew Trotter b, Þórhalla Guðmundsdóttir Beck a, Curt Anderson b, Joan Maling c, Karthik Durvasula b, Alan Beretta b,⇑ a

University of Iceland, Reykjavík, Iceland Michigan State University, East Lansing, MI, USA c Brandeis University, Waltham, MA, USA b

a r t i c l e

i n f o

Article history: Accepted 24 July 2014

Keywords: ERP Coercion Compounds Gender Icelandic

a b s t r a c t An ERP study investigated the processing of mass nouns used to convey ‘portions’ vs. ‘sorts’ interpretations in Icelandic. The sorts interpretation requires semantic Coercion to a count noun; the portions interpretation entails extra syntactic processing. Compared to a Neutral condition, Coercion escaped the expected penalty (N400), but the Extra Syntax condition incurred the anticipated costs (anterior negativity followed by P600). Furthermore, we examined the effects of having to revise an initial commitment to head-noun status. When another noun follows the mass noun (creating a compound), the second noun becomes the headnoun. We hypothesized, for Icelandic, there would be no effect for Extra Syntax because the compound should have been built before the second noun was encountered; by contrast, for the Coercion and Neutral conditions, processing costs would be incurred to detect and reconfigure the second noun as the head. These predictions were largely borne out (early and sustained anterior negativities). Ó 2014 Elsevier Inc. All rights reserved.

1. Introduction Theories that address the distribution of labor between syntax, semantics and the lexicon have sought to account for the limitless expressive power of language by adopting a principle of compositionality (e.g., Montague, 1970). This principle states that the interpretation of any structured expression is predictable from the meanings of words and the rules that combine them. For instance, the interpretation of a constituent such as black cup is a function of the meaning of black, the meaning of cup, and the syntactic rule of adjectival modification. Simple compositionality runs into numerous difficulties which provide some of the central challenges for semantics (Ramchand, 2008; Wood, 2012). This paper addresses one such difficulty: the use of mass nouns, which typically refer to unbounded substances, as count nouns, which typically refer to bounded individuals. (i) I like coffee. [mass noun referring to the substance coffee] (ii) I’ll have another coffee. [count noun referring to an implicit portion, typically a cup, of coffee] ⇑ Corresponding author. Address: Department of Linguistics, B465 Wells Hall, 619 Red Cedar Road, Michigan State University, East Lansing, MI 48824, USA. E-mail addresses: [email protected] (M. Whelpton), [email protected] (D. Trotter), [email protected] (Þ.G. Beck), [email protected] (C. Anderson), [email protected] (J. Maling), [email protected] (K. Durvasula), [email protected] (A. Beretta). http://dx.doi.org/10.1016/j.bandl.2014.07.008 0093-934X/Ó 2014 Elsevier Inc. All rights reserved.

Wiese and Maling (2005), hereafter WM, suggest that languages differ in their strategies for handling mass-to-count coercions and in the ways that the distribution of labor between syntax and semantics occurs. In particular, they suggest that, whereas cases like (ii) in English really do involve coercion (the use of a mass noun as a count noun), the equivalent example in Icelandic does not. Rather, in Icelandic there is implicit syntactic structure: the mass noun continues to be a mass noun and the countable reading is introduced by an elided head noun. The pre-theoretical domain of coercion in their account is therefore split into examples of genuine coercion (syntax–semantics mismatch) and examples of complex syntax. In this paper we investigate their analysis of the minimal contrast in Icelandic between coercion and complex syntax, using experimental data from EEG. 1.1. Mass-count coercion There is a traditional account of the mass/count distinction. Mass terms, such as water, oil, and snow, are non-pluralizable and uncountable. They apply to substances. By contrast, count terms, such as cat, flower, and bottle are pluralizable and countable. They apply to objects. In a standard view of what is involved in the mass/count division work (following Pelletier, 2010), nouns are marked with the syntactic features +mass or +count and these features serve to constrain what combinations are permissible. Combining a

45

M. Whelpton et al. / Brain & Language 136 (2014) 44–57

+mass noun oil with plural syntax or with a numeral or determiner that assumes countability would be forbidden, so two oils should be impossible, but allowable measure terms can combine straightforwardly with the singular form, e.g., more oil. Combining a +count noun bottle with plural syntax or with a numeral or determiner that assumes countability is straightforward, e.g., two bottles, but certain other quantifiers are not permitted with singular count nouns, e.g., *more bottle. The meanings of simpler units are combined in syntactically constrained ways to arrive at meanings of more complex syntactic units. However, as many have observed, some nouns can have both mass and count readings, as in (1) and (2): (1) (2)

a. b. a. b.

Turkish coffee is strong Two coffees, please Jane watched a lamb that had lost its mother Jane had lamb for dinner

To interpret the (b) examples, it has been claimed that enriched composition (Jackendoff, 1997) is required to coerce the intended meanings. In that spirit, three universal semantic coercion operations have been proposed (Pelletier, 1975), (though how universal they really are is an open question). First, a Universal Grinder which can ‘grind’ an object. Insert a count term into the grinder, say, a table, and the result is table all over the floor. So, the count noun lamb in (2a) goes in, and what comes out is lamb as a mass noun that is interpreted as food (2b). A Universal Packager applies to (1). The Packager can take a mass term, such as coffee (1a) and convert it to a count term in which portions are counted (1b). We may also think of a Universal Sorter which, like the Packager, takes mass terms as its input, as in (3a) and converts them into count terms that are interpreted as taxonomic sorts, as in (3b): (3)

a. b.

Not everyone loves wine with their dinner There are many fine Greek wines

The only study (to our knowledge) that directly examines the mass/count distinction from the perspective of the semantic coercion operations is Frisson and Frazier (2005). They report two eyetracking experiments of English, one that examines the processing costs of the Universal Packager and one that examines the processing costs of the Universal Grinder. They assumed, following Copestake and Briscoe (1995), that one sense was derived from a more basic sense via a lexical rule. Thus, an underived mass noun can yield a derived portion interpretation (via a Packer lexical rule) and, conversely, an underived count noun can yield a derived substance interpretation (via a Grinder lexical rule). Derived senses ought to have a processing cost and Frisson & Frazier found that there was indeed a cost for the derived senses in both experiments. 1.2. The present study: sorts and portions in Icelandic Different languages handle portioning and sorting in different ways. We have already alluded to the uncertainty of the divide between the two in English, but in German the division is sharply drawn (though not obligatory). Wiese and Maling (2005, p. 29) provide the following minimal pair: (4)

a. b.

zwei Bier two beer-sg. zwei Biere two beers-pl.

portion interpretation

would be suitable for restaurant orders. WM also analyzed the way that Icelandic dichotomizes portion and sorts construals, providing the theoretical background for the study we report here. German uses the plural to mark the sorts reading but Icelandic uses gender agreement to accomplish the same ends. WM proposed that a sorts interpretation of the mass noun coffee occurs in (5) but in an almost identical sentence (6), a portion interpretation obtains (like German, the distinction is not obligatory in all contexts). The difference is that in (5) the determiner another agrees in gender (neuter) with the mass noun coffee, but they do not agree in (6) where another is masculine and coffee is neuter. (5)

(6)

Get

ég

fengið

annað

kaffi

may

I

have

Get

ég

fengið

another (neut.) annan

coffee (neut.) kaffi

May

I

have

another (masc.)

coffee (neut.)

sorts interpretation

portion interpretation

The gender agreement of a determiner such as another with the nominal it modifies, coffee, would be expected in languages in which gender agreement is morphologically marked. However, in Icelandic, strikingly, the sentence is interpreted as a request for another sort of coffee (as we see in 5), not as a request for another cup (of the same kind) of coffee.1 To arrive at that sorts interpretation (in 5), it is necessary to do some extra semantic work, that is, to coerce the mass noun coffee into a countable object, a sort. In arguing for this, WM regard coercion in a way that is familiar in the foregoing discussion, namely, as ‘‘an enrichment of the semantic representation’’ (p. 17), not some extra syntactic work. As they point out, if there were a silent syntactic component in (5), it would be necessary to postulate some unexpressed element in the syntactic representation that provides the sorter interpretation, perhaps an unexpressed noun, and we would expect the gender marking to reflect that. In such a case, the numeral would agree, not with the nominal coffee, but with the silent sorter noun. The default noun for ‘sort’ or ‘kind’ in Icelandic is tegund, which is feminine, or perhaps gerð, which is also feminine, and yet the determiner in (5) is neuter, agreeing with the beverage nominal coffee. So, the gender facts do nothing to support an empty sorter noun, but everything to support semantic enrichment via coercion. By contrast, in (6) there is evidence for an unexpressed syntactic element. The interpretation of (6) is a request for another cup of (the same kind of) coffee. In (6), another and coffee do not agree in gender, but as WM argue, another agrees with the most likely container noun, in this case, cup (bolli). Because the container noun is clearly implicit, as the gender facts testify, coffee is not coerced to a portion; it remains a mass substance. The container alone is countable. Thus, the contrast between (5) and (6) is a contrast between an enriched composition, that is, a sorter coercion in the semantic representation, and an unexpressed container noun that is built in the syntactic representation. Icelandic thus offers the opportunity not only to observe the neurophysiological effects of semantic coercion to obtain a sorts interpretation, but also to view that effect in direct contrast to the effects of obtaining a portion interpretation via a silent syntactic operation. Even though (5) and (6) are minimally

sorts interpretation

The plural marker in (4b) indicates a sorts interpretation and provides a crisp contrast with the portion reading of (4a) that

1 The mass-count coercion involved here has parallels in two other kinds of coercion that have been fruitfully explored in the psycholinguistic literature, namely, complement coercion and aspectual coercion. We will have cause to discuss complement coercion elsewhere in the paper, but aspectual coercion involves coercing a telic event reading to a durative reading, as in ‘the girl hopped until dawn’ vs. ‘the girl slept until dawn’ (see Piñango, Zurif, & Jackendoff, 1999).

46


different strings, only one of them involves a mass-to-count shift and enriched semantic composition, so there ought to be a clear processing difference between the two. That is our main prediction. We also expected that sentences like (5) and (6) should both elicit a different response from a neutral sentence in which there is no silent syntactic operation and no semantic coercion, as in the following three-way comparison: (7)

Coercion:

(8)

Neutral:

(9)

Extra Syntax:

Konan þáði aðra sósu woman-the accepted another (fem.) sauce (fem.) ‘The woman accepted another sauce’ (mass > count; sort) Konan þáði aðra dúkku woman-the accepted another (fem.) doll (fem.) ‘The woman accepted another doll’ (count) Konan þáði aðra romm woman-the accepted another (fem.) rum (neut.) ‘The woman accepted another rum’ (ellipsis; portion)

either share gender with a determiner or not and still be perfectly legitimate). So, in (7), if bowl were to follow, syntactic reconfiguration would be required to change the head noun. The same would be true of (8) if doll were followed by another feminine noun, pram2 (kerru). However, for (7), in addition to the syntactic reconfiguration, there would also be a shift in meaning. If sauce is initially coerced from a mass noun to a count noun to indicate a sort of sauce, that coercion would have to be undone when bowl is encountered. To summarize, if sentences (7–9) continued into noun–noun compounds, as in (10–12), then (10) would have syntactic and semantic revisions to make on encountering bowl, (11) would have a syntactic revision to make on encountering pram, and (12) would have neither a semantic nor a syntactic revision to make on encountering bottle. (10)

(11)

(12) Sentences like (7) are generally interpreted as referring to another sort of sauce, rather than more of the same sauce. Sentences like (8) are neutral in the sense that with count-nouns, like doll, that possess the same gender agreement pattern as (7), a sorts reading is much less salient. In (9), the determiner another does not agree with the mass noun rum, but it does agree with an unexpressed noun, bottle (flösku), which has to be built into the syntax, and a portion reading follows. Assume, then, that a sentence like (9) involves building the container noun (bottle) into the syntactic representation; but there is another way that the portion reading can be rendered, that is, by simply expressing the container noun in what would be a noun–noun compound: rum bottle (which in Icelandic, unlike in English, it is worth noting, is capable of meaning either a bottle for rum or a bottle of rum). Because nothing in the morphology of many Icelandic nouns gives a clue that a compound is imminent, a comprehender encounters rum first and, if the container noun is immediately built into the syntactic representation, then there should be no syntactic work left for the processor to do when the container noun bottle appears (assuming pragmatic contexts which demand the of rather than the for interpretation, which was the case in our experimental sentences). By contrast, in sentences such as (7) and (8), if a container noun were to follow sauce and doll, there would be syntactic work to do. In (9), if the container noun is built independently of the compound expression, then there can be no commitment to rum as the head noun of the compound, but in (7) and (8), such a commitment ought to be made given that there is no morphological clue to indicate otherwise. In (7), the coerced reading of a sort is only possible if sauce is taken to be the head noun (the head of a compound determines the category of the compound; so, an adjective + noun compound, such as blackboard, would be a noun; in Germanic languages, including Icelandic, the rightmost element of a compound is always the head). That initial commitment to sauce as the head would have to be revised if a suitable container such as bowl (skál) that agrees in gender with another were to follow. Bowl would then become the head noun, and the earlier assumption that sauce agreed with another would be revised to an incidental sharing of gender but not syntactic agreement (the first noun of a noun–noun compound in other Germanic languages, too, can

Konan þáði aðra sósu skál woman-the accepted another (fem.) sauce (fem.) bowl (fem.) ‘The woman accepted another sauce bowl’ Konan þáði aðra dúkku kerru woman-the accepted another (fem.) doll (fem.) pram (fem.) ‘The woman accepted another doll pram’ Konan þáði aðra romm flösku woman-the accepted another (fem.) rum (neut.) bottle (fem.) ‘The woman accepted another rum bottle’

The potential demands made of the parser at the first constituent of the compound (C1) and the second constituent of the compound (C2) are summarized in Table 1. In a study designed to ensure that subjects never know if the sentence will end in a single noun or a noun–noun compound and mostly expect the former, it ought to be possible to observe the processing costs incurred by coercion at the single noun (which we will henceforth refer to as C1 for constituent one) in (10), the costs of building silent syntax at C1 (in 12); and at C2 (constituent two), the differential effects of having to revise or not commitments made at C1. That is what the present study set out to do. 1.3. Expected electrophysiological effects What effects might we expect in an ERP experiment, first, for the mass-count coercion at C1 in (10) and, second, for the silent syntactic operation at C1 in (12)? Regarding coercion, since it is a semantic process, an N400 might be the expected electrophysiological effect, although until very recently, there had been no report of an N400 response, or in fact any ERP response, to coercion. Fortunately, two recent reports of ERP responses to complement coercion provide guidance. Kuperberg et al. (2010) and Baggio et al. (2009) both conducted almost identical ERP studies of complement coercion, so their findings ought to be instructive regarding what we might expect. Both studies, although they found differing later negativities, found an effect in common that they characterized as an N400 at the noun phrase that is coerced from an entity to an event interpretation relative to a non-coerced condition, a cost that is consistent with integrating an enriched meaning into the semantic representation.3 That is the main finding and it indicates that the 2 The word pram is perhaps not familiar to all American speakers, who may prefer the term stroller. 3 The N400 is also thought to be consistent with lexical access. For the purposes of the present paper, we remain agnostic about whether the N400 indexes lexical access or semantic integration, even if we use the language of semantic integration here.

M. Whelpton et al. / Brain & Language 136 (2014) 44–57 Table 1 The different demands that each condition makes at C1 and C2. Condition

C1

C2

Coercion

mass-count coercion; assume C1 is head-N assume C1 is head-N gender mismatch; build unexpressed-N into syntax; assume it is head-N

revise coercion, and head-N to C2 revise head-N to C2 Unexpressed-N now expressed; no revision

Neutral Extra Syntax

same effect might be anticipated at the nominal sauce in (10) where an enriched interpretation has also to be integrated into the semantic representation. The only caution is that a recent experiment by Delogu, Crocker, and Drenhaus (2012) has raised questions about whether the effects reported in these studies are really due to semantic mismatch or to degree of surprisal. So, it will be important to rule surprisal out of the equation to avoid a potential confound (see discussion of the cloze pretest in the Methods section). Turning to the effects of a silent syntactic operation, no study that we are aware of has investigated the effects of building an unexpressed container noun in the syntactic representation, so there is no direct precedent that is instructive. However, it is possible to make reasonable inferences about what sorts of effect might be anticipated. First of all, many syntactic violations have elicited a P600 or a LAN (Friederici, 1995, 2002; Gouvea, Phillips, Kazanina, & Poeppel, 2010; Hagoort, Wassenaar, & Brown, 2003), so in general terms, these would seem the most likely responses to sentences such as (12) at C1. More specifically, recall that in (12), there is a mismatch in gender between the determiner and the nominal; violations of gender have given rise to a P600 (e.g., Allen, Badecker, & Osterhout, 2003; Coulson, King, & Kutas, 1998; Gunter, Stowe, & Mulder, 1997; Hagoort et al., 2003; Osterhout & Mobley, 1995) but in some studies, a LAN and a P600 (e.g., Barber, Salillas, & Carreiras, 2004; Deutsch & Bentin, 2001; Gunter, Friederici, & Schriefers, 2000; Neville, Nicol, Barss, Forster, & Garrett, 1991; Rösler, Pütz, Friederici, & Hahne, 1993), so whether the observed effect for (12) is a LAN or a P600 or both, it should certainly contrast with the expected N400 to (10). Let us examine more closely what we might expect for the silent syntactic work involved in (12) by homing in on studies that have the closest parallels with our own in the relevant respects. Most studies that have investigated gender in ERP studies have made use of gender violations. Sentence (12) involves a mismatch in grammatical gender between the determiner and C1, but there is no violation. If the sentence ended at C1, it would be perfectly fine, as provision is immediately made for the unexpressed C2 that does agree with the determiner. For a case like this, the compound literature is especially illuminating as several studies have addressed the effects of gender incongruities at C1, even though they are not germane to the well-formedness of the compound, where only C2 has to agree with the determiner, but C1 may incidentally match or not. In a study of gender agreement between determiners and compounds in German, Koester, Gunter, Wagner, and Friederici (2004) and Koester, Gunter, and Wagner (2007) manipulated gender at C1 and reported a LAN for incongruities. In Koester et al. (2004), two of their four conditions involved a mismatch in gender between the determiner and C1: (13) (14)

Das The (neut.) * Das The (neut.)

Presse press (fem.) Nuss nut (fem.)

amt office (neut.) baum tree (masc.)

47

In (13), there is no violation. The determiner the (das) happens not to match the gender of the C1 noun press (Presse), but it does agree with the C2 noun office (amt). In (14), as indicated by the asterisk, there is a violation; the determiner incidentally happens not to match the gender of C1, but crucially does not agree with C2 (which is where the violation lies). But it cannot be known at C1 that the violation at C2 is going to occur. In both (13) and (14), all that the parser could notice at C1 is that there happens to be a gender mismatch, but that mismatch is irrelevant to the grammaticality of the compound. Nevertheless, registering the mismatch in (13) and (14) does incur a processing cost. Koester et al., report a LAN for both such cases. We note that while many studies mentioned above report a P600 for gender mismatches, Koester et al. (2004) do not; they obtained a LAN, but no P600, and theirs is the only compound study among them. (We will take this up in the discussion section.) At C2, predictions are a little more speculative, as there is very little in the literature to go on. For Coercion, if an N400 is observed at C1, then perhaps an N400 should be observed at C2 as well as the coercion is undone. But the Coercion condition, like the Neutral condition, also has to shift head-noun status from C1 to C2, so there should be some syntactic cost there. This could be a (E)LAN, an early automatic response to structure building (Friederici, 2002); but it could also be followed by a P600 as the head-noun status is revised; Koester, Holle, and Gunter (2009) report a P600 in three-constituent compounds when a C2 has to renounce its head-noun commitment as a C3 is heard. Regarding the Extra Syntax condition, by comparison with the other two conditions, the ERP signal should be muted since C2 for this condition is redundant. 2. Methods 2.1. Participants 23 right-handed native speakers of Icelandic from the University of Iceland (15 women, 8 men) with a mean age of 25.3 years (range, 20–54) took part in the experiment. Subjects had normal or corrected-to-normal visual acuity. They gave their informed consent and were paid for their participation. Since an excessive number of trials for 5 subjects were contaminated by artifacts, 18 subjects were included in the final analysis (13 women, 5 men) with a mean age of 24.5 years (range 20–33). 2.2. Stimuli The set of experimental stimulus materials consisted of three critical conditions with 35 sentences per condition (see Table 2; from sentences 10–12 reproduced here for convenience; the determiner and the compound in the table follow a common sentence beginning, in this case, Konan þáði. . . (‘the woman accepted. . .’): (i) mass-noun coercion, gender match between Determiner, C1 and C2, (ii) neutral, count-noun, gender match between Determiner, C1 and C2, and (iii) mass-noun, gender mismatch between determiner and C1. Each triplet of experimental sentences was identical except for the compound noun in object position. The noun–noun compound was presented as separate words. In Icelandic, the written form of compounds is always as a single word, but we presented each noun separately so as to garden-path participants into assuming that the first noun encountered was most likely a single noun. To offset any possible initial surprise at this rendering of compounds, all compounds used in practice items and in subject position in fillers were also presented as separate words. Participants reported that they quickly got used to this mode of presentation and that it was barely noticeable.

48


Table 2 Example sentences for each of the three critical conditions. Condition

Example

Coercion

aðra another (fem.)

sósu sauce (fem.)

skál bowl (fem)

Neutral


dúkku doll (fem.)

kerru pram (fem.)

Extra syntax


romm rum (neut.)

flösku bottle (fem.)

Nouns at C1 in the experimental items were chosen carefully. A key factor was that the morphology of the nouns as single nouns and as the first element in a noun–noun compound had to be identical. It was necessary for there to be no cue to participants that a compound noun would occur (a goal that was furthered by all 280 fillers containing a single noun in object position; see below). Frequencies of the N–N compounds in the experimental sentences were checked to the extent possible. Icelandic is considered a less-resourced language in terms of language technology. According to Daðason, ‘‘In META-NET’s language report for Icelandic (. . .), the language was categorized as having weak or no support for language resources’’ (2012, p. 12). However, impressive efforts have been made to remedy the situation and we were able to make use of a 25-million word lemmatized corpus known as the Tagged Icelandic Corpus (Helgadóttir, Svavarsdóttir, Rögnvaldsson, Bjarnadóttir, & Loftsson, 2012), which, in spite of whatever lacunae it may possess, at least provided some basis for controlling frequencies in our stimuli. Compound frequencies were not significantly different between the three experimental conditions. Coercion (mean 7.09) vs. Extra Syntax (mean 9.21), t(34) = 0.77, ns. Coercion (mean 7.09) vs. Neutral (mean 3.38), t(34) = 0.15, ns. Neutral (mean 3.38) vs. Extra Syntax (mean 9.21), t(34) = 0.40, ns. At C1, the only significant difference was that Neutral was more frequent than Extra Syntax (p < .05); at C2, there were no significant differences. However, interpreting these frequencies requires caution, as there were many words that had zero occurrences in the corpus we used even though they are considered common in everyday speech; words such as cappuccino or expresso barely registered at all in spite of their daily familiarity in cafeteria and restaurant talk. Three pretests on experimental stimuli were conducted, a cloze test, a plausibility rating task, and a portion-sorts rating task. The Cloze test was included because we needed to be sure that any effect for the Coercion condition at C1 was not due to its having a lower cloze probability than the other conditions. 34 participants were asked to supply a word to follow the 35 sentence frames that were common to all three experimental conditions (in Table 1, this would be: ‘‘the woman accepted ______’’). Given that the sentences did not appear to constrain the choice of critical word, it seemed unlikely that one condition would be disadvantaged, and the cloze test confirmed this. Only three words that even a single respondent offered got a single match with any word in the experimental stimuli (one for each condition). In the plausibility rating task, subjects were asked to indicate on a 5-point scale whether or not a sentence made sense (‘1’ for ‘makes no sense’ to ‘5’ for ‘makes total sense’). The task included all sentences from the three experimental conditions and the two anomalous sets of fillers. The anomalous fillers were judged to be highly implausible (mean for incongruous meaning, 1.46; mean for number violations, 1.42). By comparison, each experimental condition was considered highly plausible (means of 3.77, 3.79, and 4.41 for Coercion, Neutral, and Extra Syntax conditions, respectively). An ANOVA comparing the means the three experimental conditions was significant [F(2, 102) = 6.33, p < 0.05]. HSD

post-tests showed that while the Coercion and Neutral conditions were not significantly different from each other, they were both rated less plausible than the Extra Syntax condition, p < 0.01 in both cases. The portion-sorts rating task was to establish that stimuli for the Extra Syntax (portions) and Coercion (sorts) conditions were interpreted as intended. Subjects were presented with all of the stimuli from these two conditions and asked to rate on a 5-point scale whether the sequence ‘another + noun’ meant more of the same thing (portions) or a different thing (sorts), ‘1’ for ‘same thing’ and ‘5’ for ‘different thing’. As expected, the means for each condition (2.04 for the Extra syntax portions sentences, 3.74 for the Coercion sorts sentences) were significantly different: t(34) = 10.8227, p < 0.0001. The mass and count nouns in the three experimental conditions were controlled for length (2 syllables per word on average at C1, 1.6 syllables per word on average at C2). There were no significant differences between conditions at either C1 or C2. In addition to the 105 experimental stimuli, 280 fillers were included, including 35 sentences with incongruous nouns at the same position in the sentence as C1 nouns in the experimental stimuli, and 35 sentences with violations of number agreement between the determiner and the noun, again at the same position in the sentence as the C1 nouns in the experimental stimuli. The inclusion of these anomalous fillers served to keep participants off-balance by making it uncertain what sort of noun might occur at the critical part of the sentence. The fillers also contained compound nouns in subject position (presented as separate words) and half of the fillers contained the sequence another N in subject position to prevent participants from expecting that such a sequence would only occur in object position (as it does in the experimental stimuli). All 280 fillers had single nouns in object position so as to provide an overwhelming expectation that a single noun was to be expected there, reinforcing the likelihood that participants would be garden-pathed in the experimental sentences and would thus not be expecting a compound to occur in object position. 2.3. Procedure Participants sat in a dimly-lit, sound-attenuated, and electrically-shielded room separate from the experimenter. They were instructed to pay attention to the sentence as it appeared on the screen in front of them a word at a time, and to judge whether a word (in red letters) that followed the sentence was related to the sentence or not. For example, if the sentence were The waiter recommended a glass of cognac, then the word whiskey would be related but the word hammer would not. Y/N responses were indicated via one of two keystrokes (counterbalanced across participants). Subjects were given 5 practice trials before the experiment began. We chose this task, even though it can be quite demanding, as it focused subjects’ attention on meaning, which is clearly relevant to this study. Sentences were presented word-by-word on a computer screen. Each trial began with a fixation cross at the center of the screen for 800 ms, followed by a 200-ms blank screen, followed by the first word of the sentence. Each word appeared on the screen for 400 ms with an ISI of 200 ms between each word. The final word of each sentence was followed by a 1200-ms blank screen interval and then the task word appeared in red letters at the center of the screen. The task word remained on the screen until the subjects made a response or for a maximum of 3 s. At this point, the next trial started. The session lasted for approximately 50 min. The 385 sentences were divided into 8 blocks of 50 trials. The order of presentation was randomized for each subject.


49

2.4. Recording

3.2. ERP data: C1

The electroencephalogram (EEG) was recorded from 32 sintered Ag/AgCl electrodes held in place on the scalp by an elastic cap (GND WaveGuard 64 Electrode cap; Advanced Neuro Technology BV., Enschede, The Netherlands). It was amplified using a Full-band EEG DC Amplifier (Advanced Neuro Technology), with a 256 Hz sampling rate (and a bandpass filter of 0.01–40 Hz which was applied offline). Electrode impedances were kept below 10 kX, mostly below 5 kX (active shielding in the caps permits clean data acquisition even when skin impedances are higher than conventional standards). Averaged mastoids were used as reference. All signals were recorded continuously.

The mean ERP amplitude was calculated for each of the three conditions and for each region of interest (ROI) in the time windows 350–450 ms and 500–700 ms time-locked to the onset of C1 (Fig. 2). An ANOVA was performed with the factors ROI (7: AC, AL, AR, CC, Cl, CR, PC) and condition (3: Coercion, Neutral, Extra Syntax). At the 350–450 ms time window, there was a main effect of ROI [F(6, 102) = 5.37, MSE = 7.097, p = 0.0069]. There was no main effect for condition [F(2, 34) = 0.31, p = 0.70]. There was an ROI by Condition interaction [F(12, 204) = 4.71, MSE = 1.613, p = 0.0013]. We conducted one-way ANOVAs to determine where the interactions took place. These were significant only at two ROIs, AC and AR. At AC, [F(2, 34) = 4.92, MSE = 4.99, p = 0.02]; at AR, [F(2, 34) = 3.64, MSE = 4.024, p = 0.04]. Follow-up t-tests revealed that at both ROIs, the Extra Syntax condition was significantly more negative than both Coercion and Neutral conditions (see Fig. 3); at AC, Extra Syntax compared with Coercion, t(17) = 2.40, p = 0.028, and compared with Neutral, t(17) = 2.59, p = 0.019; at AR, Extra Syntax compared with Coercion, t(17) = 2.14, p = 0.047, and compared with Neutral, t(17) = 2.25, p = 0.038. There was no difference between the Coercion and Neutral conditions at either ROI. At the 500–700 ms time window, there was a main effect for ROI, [F(6, 102) = 4.91, MSE = 7.059, p = 0.015] and for Condition, [F(2, 34) = 4.75, MSE = 18.559, p = 0.015]. There was also an interaction of ROI by Condition, [F(12, 204) = 7.92, MSE = 1.440, p = 0.00001]. One-way ANOVAs were conducted to determine the scalp locations of the interactions. There were significant effects at four ROIs: CC [F(2, 34) = 5.19, MSE = 6.25, p = 0.01]; CL [F(2, 34) = 14.26, MSE = 3.44, p = 0.00003]; CR [F(2, 34) = 8.93, MSE = 3.10, p = 0.0008]; and PC [F(2, 34) = 8.75, MSE = 3.44, p = 0.0009]; (see Fig. 4). T-tests showed that at each ROI, while there was no difference between Coercion and Neutral conditions, the Extra Syntax condition was more positive than both Coercion and Neutral conditions. At CC, Extra Syntax compared with Coercion, t(17) = 3.29, p = 0.004; Extra Syntax compared with Neutral, t(17) = 2.35, p = 0.03. At CL, Extra Syntax vs. Coercion, t(17) = 5.57, p = 0.00003; Extra Syntax vs. Neutral, t(17) = 3.85, p = 0.001. At CR, Extra Syntax vs. Coercion, t(17) = 4.27, p = 0.0005; Extra Syntax vs. Neutral, t(17) = 2.99, p = 0.008. At PC, Extra Syntax vs. Coercion, t(17) = 3.46, p = 0.003; Extra Syntax vs. Neutral, t(17) = 3.92, p = 0.001. Unexpectedly, there appeared to be an early effect during the 50–150 ms time window, but while there was a main effect for ROI [F(6, 102) = 14.91, MSE = 4.29, p = 0.000009], there was no main effect for condition [F(2, 34) = 1.22, n.s.] and no interaction [F(12, 204) = 1.90, n.s.].

2.5. Data analysis Average ERPs were calculated across subjects for C1 and C2 by condition for each compound constituent separately. The averaged epoch extended from 200 ms preconstituent onset to 800 ms postconstituent onset for both C1 and C2. Seven ROIs were constructed to test the ERP for scalp distributions (see Fig. 1). The ROIs were defined as: Anterior Central (AC): Fp1, Fpz, Fp2, Fz; Anterior Left (AL): F3, F7, FC1, FC5; Anterior Right (AR): F4, F8, FC2, FC6; Centroparietal Central (CC): Cz, CP1, CP2, Pz; Centroparietal Left (CL): C3, CP5, P3, P7; Centroparietal Right (CP): C4, CP6, P4, P8; and Posterior Central (PC): POz, O1, Oz, O2. Contaminated EEG epochs were automatically rejected and double-checked by visual inspection. Artifact rejection no more stringent than ±75 lV criterion resulted in approximately 20% of the trials being excluded, similarly in both C1 and C2. The ERP waveforms were quantified by mean amplitude measures in relation to a 200-ms prestimulus baseline (immediately prior to C1 and again to C2). ROIs and conditions were subjected to repeated measures ANOVAs at C1 and at C2, using R (R Core Team, 2013). In all ANOVAs, in order to guard against Type 1 error due to violations of sphericity, the Greenhouse–Geisser correction was applied where appropriate. In all of these cases, we reported the original degrees of freedom with the corrected p value. 3. Results 3.1. Behavioral data Subjects reported no difficulty and performed adequately in the task given that it was intentionally a reasonably demanding one. The accuracy rate was 72% and indicates that participants were following instructions.

Fig. 1. Electrode layout and ROIs. AL (Anterior Left); AC (Anterior Central); AR (Anterior Right); CPL (Centroparietal Left); CPC (Centroparietal Central); CPR (Centroparietal Right); PC (Posterior Central).

50


Fig. 2. ERPs for the first compound constituent (C1) for the three conditions: Coercion (red), Extra Syntax (green), and Neutral (blue). Significant differences are shaded. Negative is plotted upward.

Fig. 3. Anterior Negativity (350–450 ms) at C1.

3.3. ERP data: C2 As with C1, the mean ERP amplitude was calculated for each of the three conditions and for each region of interest (ROI) in the 150–350 ms and 500–700 ms time-windows locked to the onset of C2 (Fig. 5). An ANOVA was performed with the factors ROI (7: AC, AL, AR, CC, Cl, CR, PC) and condition (3: Coercion, Neutral, Extra Syntax). At the 150–350 time window, there was a main effect for ROI [F(6, 102) = 3.48, MSE = 6.80, p = 0.049] and for condition [F(2, 34) = 4.91, MSE = 31.74, p = 0.013]. There was an ROI by condition interaction [F(12, 204) = 3.09, MSE = 4.60, p = 0.049].

One-way ANOVAs revealed significant differences at AL [F(2, 34) = 3.30, MSE = 9.36, p = 0.049], AC [F(2, 34) = 4.67, MSE = 32.38, p = 0.017], AR [F(2, 34) = 5.25, MSE = 4.91, p = 0.019], and CC [F(2, 34) = 3.45, MSE = 5.11, p = 0.043]; see Fig. 6. At these ROIs, Coercion and Neutral were not significantly different from each other, but both conditions were in most cases significantly more negative than Extra Syntax. At AL, Coercion vs. Extra Syntax, t(17) = 1.22, n.s.; Neutral vs. Extra Syntax, t(17) = 2.62, p = 0.018). At AC, Coercion vs. Extra Syntax, t(17) = 1.96, p = 0.066 (marginally significant); Neutral vs. Extra Syntax, t(17) = 3.13, p = 0.006. At AR, Coercion vs. Extra Syntax, t(17) = 2.04, p = 0.057 (marginally significant); Neutral vs. Extra


51

Fig. 4. Late positivity (500–700 ms) at C1.

Syntax, t(17) = 2.86, p = 0.011. (At this early time window, no comparisons were made at centroparietal or posterior ROIs because of an expected component overlap from the observed late positivity in those regions at C1, so no further analyses were conducted on CC. See Section 4.) After visual inspection of the waveforms, additional analyses were carried out during the time window 350–500 ms. There was no main effect for condition, but there was a main effect of ROI [F(6, 102) = 7.31, MSE = 7.86, p = 0.0015] and an ROI by condition interaction [F(12, 204) = 3.64, MSE = 5.78, p = 0.044]. One-way ANOVAs determined that there were effects at AC [F(2, 34) = 3.54, MSE = 41.86, p = 0.059 (marginal)] and AR [F(2, 34) = 6.016, MSE = 4.00, p = 0.008]. Two-tailed t-tests indicated that at these ROIs, Coercion and Neutral were not different

from each other, but both were significantly more negative than Extra Syntax. At AC, Coercion vs. Extra Syntax, t(17) = 2.245, p = 0.038; and Neutral vs. Extra Syntax, t(17) = 2.29, p = 0.035. At AR, Coercion vs. Extra Syntax, t(17) = 2.4, p = 0.028; Neutral vs. Extra Syntax, t(17) = 2.97, p = 0.008 (see Fig. 7). During the 500–700 ms time window, there was a main effect for ROI [F(6, 102) = 10.44, p < 0.03]. There was no main effect for condition, and no interaction.

4. Discussion The present article investigated the effects of semantic coercion and the effects of building an unexpressed element into the syntax.

Fig. 5. ERPs for the second compound constituent (C2) for the three conditions: Coercion (red), Extra Syntax (green), and Neutral (blue). Significant differences are shaded.

52



Coercion is thought to be non-compositional and in processing terms has been associated with extra work for the parser to handle; but even the compositional processes of the combinatorial syntax can impose an increased burden on the parser. Several findings emerge from the results. Most generally, the ERP profiles for the Coercion and Extra Syntax conditions were totally different throughout C1 and C2. Plainly, very different processes are involved in these superficially minimally different sequences. More specifically, several findings are quite striking. First, there was no significant effect for Coercion at C1. Second there was a very clear effect for Extra Syntax at C1. Third, at C2, the Coercion and Neutral conditions both patterned together and were distinct from Extra Syntax. Let us deal with each in turn. 4.1. No N400 for Coercion at C1 First the lack of an effect for Coercion. For the three centroparietal ROIs between 300 and 500 ms (the usual N400 time window), the nonsignificant p values were very far from significance: Centroparietal Left: p = 0.83; Centroparietal Right: p = 0.61; Centroparietal Central: p = 0.82. This was something of a surprise as we had fully expected, on the basis of two recent studies (Baggio, Choma, van Lambalgen, & Hagoort, 2009; Kuperberg, Choi, Cohn, Paczynski, & Jackendoff, 2010) to observe an N400 for the coercion at C1, as they had done for complement coercion. We now consider why this might be. An interesting candidate explanation, alluded to earlier, for the lack of an ERP effect for Coercion, is that once surprisal is controlled for, coercion costs vanish. Delogu et al. (2012) hypothesized that the N400 response reported for complement coercion might in

fact be a response to surprisal (the inverse probability of a word’s occurrence in a context). We wished to make sure that if we were to obtain an effect for Coercion that it would not be subject to a surprisal construal. A cloze test established that there was no difference between conditions for the probability of occurrence of the noun at C1. So, according to the definition of surprisal Delogu et al. were guided by, our stimuli could not be interpreted favoring one condition over another in terms of what words were expected at C1. And perhaps that is the reason we did not obtain an effect. No surprise, no N400. A further possible reason for the failure of Coercion to elicit an ERP response is suggested by Frisson and Frazier (2005). In their experiment examining mass-to-count coercion, they compared a neutral condition involving a bare mass noun pluralized (beers) with a ‘helping’ condition in which a numeral preceded the coerced nominal (three beers). The presence of the numeral signaled that whatever the nominal was going to be, it was going to be countable, and they found that this advance warning helped make processing much easier than in the neutral condition. In our experimental stimuli, the determiner another preceded the coerced nominal; this determiner may have functioned in the same way as Frisson & Frazier’s numeral, signaling that the nominal is going to be countable, and thus ameliorating possible processing difficulties. A related explanation suggested by an anonymous reviewer concerns the hypothesis that nouns are semantically underspecified for the distinction between sorts and portions interpretations and that what determines their count or mass status is the syntax in which they are embedded, a view that is defended in Borer (2005). If this view is correct, then the absence of an N400 follows



(though it would render the Frisson & Frazier effect for mass-count coercion mysterious). 4.2. Anterior negativities and a later positivity for Extra Syntax at C1 While Coercion did not yield a noticeable ERP response at C1, the Extra Syntax condition played out almost as predicted. We found an early (350–450 ms) negativity in anterior central and anterior right ROIs for Extra Syntax relative to both Coercion and Neutral. This was followed by a later positivity (500–700 ms) in four centroparietal and posterior ROIs, again for Extra Syntax relative to both Coercion and Neutral. The scalp distribution of the anterior negativity at 350–450 ms is at central and right ROIs, so it is not identical to the left-dominant effect (LAN) that has often been reported for gender mismatches. But there are precedents for more distributed anterior negativities to morphological mismatches. Münte and Heinze (1994) found anterior negativities to gender violations in German and Finnish that were not confined to the left but were more distributed. They point to similarities with LAN. See also Vos, Gunter, Kolk, and Mulder (2001) for a review noting the broad variety of anterior negativities in a range of studies. There are also direct precedents for a right anterior negativity, a RAN, in response to morphological mismatches. A RAN was obtained for illicit morphology in two experiments investigating wh-questions in Japanese (Ueno & Kluender, 2009). They reported a RAN between wh-question words and Q(uestion) morpheme. Discussing these findings, they argue that although similar studies in English and German elicited a LAN, their RAN is probably closely related, and they use the formulation ‘‘LAN/RAN’’ to convey this. They point out that when different languages are investigated, differences in syntax ‘‘may trigger the same, different, or partially overlapping processes, resulting in the same, different, or partially overlapping ERP effects . . . which can lead to a better understanding of anterior negativities’’ (2009, p. 66). Confirming this, a recent ERP study of Hindi also elicited a clear RAN in response to anomalous morphology (Dillon, Nevins, Austin, & Phillips, 2012). Like Münte and Heinze (1994), Ueno and Kluender (2009), and Dillon et al. (2012), we take it that the LAN and our central and right anterior negativities are related, syntactically-driven responses associated with morphology (notably in the case of Münte & Heinze, gender morphology) and proceed with our discussion. We interpret the anterior negativities in our results in the same way as Koester et al. (2004) interpreted the LAN that they obtained for a gender mismatch at C1, that is, as ‘‘an automatic checking of a gender agreement of two syntactic elements’’ (p. 1652). A number of other studies have also viewed anterior negativities as a reflection of the detection of morphosyntactic mismatches (e.g., Coulson et al., 1998; Weyerts, Penke, Dohrn, Clahsen, & Münte, 1997) and the assumption that detecting gender mismatches is an automatic process, one that is not subject to strategic control, is widely accepted (e.g., Gunter et al., 2000). Interpreting the anterior negativity at 350–450 ms for Extra Syntax relative to Coercion and Neutral in this way is perhaps rather straightforward, given the ample parallels in the literature, but Koester et al. (2004) express some puzzlement regarding the purpose of gender-checking, and in attempting to solve this puzzle we seek also to make sense of another puzzle, namely, that we obtained a P600 at C1 but Koester et al. did not. They propose that, if checking for gender agreement is what triggered the LAN response, ‘‘this might help to determine that a constituent is not the last’’ and that ‘‘a further constituent has to be processed’’, but the problem they see is that ‘‘such a mechanism cannot fully suffice because nonhead constituents may also agree in gender with a determiner. Thus the function reflected by the nonhead LAN

53

cannot be fully (e.g., Friederici, 1995) clarified with the present data’’ (Koester et al., 2004, p. 1652). So, noticing a gender mismatch between determiner and C1 serves to prepare the parser for a C2, but if there happens to be a gender match at C1, then it can hardly serve the same purpose because the nonhead noun could be the head noun and the subsequent arrival of C2 should then be unexpected. This seems to be inevitable, as we already have good evidence that the parser will commit to a head-noun analysis of the nonhead noun in the absence of any morphological information to preclude that commitment. Such evidence is provided by Staub, Rayner, Pollatsek, Hyönä, and Majewski (2007) who compared reading times for sentences in which the nonhead noun of a noun–noun compound was plausible as the head noun with sentences in which the nonhead noun was implausible as the head noun: (15) (16)

The new principal visited the cafeteria manager The new principal talked to the cafeteria manager

The nonhead noun cafeteria led to inflated reading times (including the very early first fixation times) where it is implausibly the head noun, as in (16), than where it is plausibly the head noun, as in (15). ‘‘These findings suggest (. . .) that the parser initially analyzes a singular noun as a head instead of a modifier’’ (Staub et al., 2007, p. 1162). We suggest that the function reflected by the LAN in Koester et al. and the broader anterior negativity in the present study is only to recognize the gender mismatch at C1 and thus to tip off the parser that C1 cannot be the head noun. At this point, what happens in German and in Icelandic diverge. In German, nothing further can be accomplished at C1 because there is no proposal that an unexpressed C2 is built into the syntax, so the parser should expect an expressed C2 to follow, thus lightening the processing burden at C2. When there is a gender match at C1 in German, because the parser would initially analyze the nonhead noun as the head noun, there should be no anterior negativity/LAN at C1 (which is what Koester et al. report), but there should be an effect of revision of head-noun status at C2. For Koester et al., this does not occur; they reported a possible P600 at C2, but only for gender violations at C2, and they point out that this is probably uninterpretable because the task cue was presented at 200 ms post C2 offset, and so there would be many possible processes confounding the issue. Using German for contrast, let us work through the likely processing steps for Icelandic following the early anterior negativity for the nonhead noun in the Extra Syntax condition. With the mismatch having been detected, the parser builds the most likely unexpressed C2 into the syntax at C1; there is now no need for a C2 to follow; if it does, it is redundant, and there should be no structural work left to do. Building the silent syntax at C1, however, should have some processing cost, and we propose that the observed positivity between 500 and 700 ms is best interpreted as reflecting that cost. On this view, the fact that such a positivity was not found in German at C1 is not surprising, since no comparable unexpressed structure is built. Given the scalp distribution (centroparietal and posterior) of the effect, its timing (500–700 ms), and the fact that syntactic processing is at issue, it is probably reasonable to assume that the positivity we observed is a P600. The suggestion that it might reflect the construction of an unexpressed syntactic element does not stretch the range of functions associated with the P600 unduly. The P600 is often considered to reflect syntactic reanalysis and repair (e.g., Friederici, 1995, 2002; Hagoort, Brown, & Osterhout, 1999). It is observed in garden-path experiments, for example, when reanalysis is required after an initial parse has failed. We

54


think that the P600 we observed could be due to the kind of reanalysis required to accommodate the extra (silent) structure. 4.3. The effects at C2: early anterior negativity At C2, during the 150–350 ms time window, we found increased anterior negativity, as predicted; this occurred at all three anterior ROIs (AL, AC, and AR). We did not find the expected late positivity at centroparietal and posterior ROIs. With only slight variations, the basic picture is that the Coercion and Neutral conditions were more negative at these ROIs than the Extra Syntax condition, but not different from each other. Before trying to interpret these results, let us first deal with a possible concern. Could the observed effects at C2 reflect end-ofsentence wrap-up effects? We think this is unlikely for two reasons. First, so far as we know, wrap-up effects have always been reported following some violation (though see Kuperberg et al., 2010), but all of our experimental sentences were perfectly acceptable sentences and judged to be highly plausible. Secondly, wrap-up effects to violations are usually associated with posterior N400-like responses (e.g., Hagoort, Brown, & Groothusen, 1993; Molinaro, Vespignani, & Job, 2008; Osterhout & Holcomb, 1992) or arguably with long-sustained and broadly distributed negativity (e.g., Friederici & Frisch, 2000), but that is not the profile we observe. For these two reasons, we conclude that the effects we report are not due to wrap-up effects, and proceed to consider what might have caused them. Also, before proceeding to interpret the results, it needs to be recognized that any early effect at C2 in the centroparietal and posterior ROIs is almost certainly due to component overlap from the large P600 at C1. As can be seen in Fig. 6, in all of the ROIs that occasioned a P600 at C1 (cf. Fig. 2), immediately after the prestimulus baseline, the waveforms diverge, with the Extra Syntax condition immediately more positive. Thus, we do not ascribe these effects to C2, but to C1 spillover. Now, let us proceed to interpret the observed results at C2. We believe the effects at C2 are due to the need to revise head-noun status or not. The premise is that the parser commits to a headnoun reading of C1 if at all plausible (following the Staub et al., 2007 finding of an early response to that effect). A head-noun reading of C1 is plausible in the Coercion and Neutral conditions. Not only is it plausible, but it was actively encouraged by having 280 fillers end in a single noun rather than a compound. A head-noun reading of C1 in the Extra Syntax condition, by contrast, is ruled out by the mismatch in gender. Therefore, for Coercion and Neutral, when C2 is encountered, restructuring is required to revise the head noun from C1 to C2. For Extra Syntax, such restructuring is not necessary, as all of the restructuring, by hypothesis, was accomplished at C1. Originally, we had expected an effect for coercion at C1 and for the undoing of the coercion at C2, but since no effect for coercion was seen at C1, none was expected for a meaning shift at C2. Therefore, at C2, Coercion and Neutral ought to behave in the same way, reacting only to the need to revise which noun of the compound is the head noun. This is what we see in the ERP waveform, Coercion and Neutral not different from each other. We also see that they are significantly different from Extra Syntax. So, the contrast is two-way: head-noun revision for Coercion and Neutral vs. no head-noun revision for Extra Syntax. We now consider how this contrast is relates to the early negativity. The early left, central, and right anterior negativity occurs in a time window (150–350 ms) often associated with an ELAN; a number of studies find an ELAN-like response bilaterally (e.g., Hahne & Friederici, 2002; Pakulak & Neville, 2010; Roehm & Haider, 2009; Yamada & Neville, 2007). The ELAN (or ERAN) is frequently considered to reflect an early automatic process of phrase structure

building, when word category information is used to build initial syntactic structure (e.g., Friederici, 1995). It is reported in sentences that feature word-category violations, such as, ‘*Max’s of proof the theorem’ (Neville et al., 1991). However, as Kutas, Van Petten, and Kluender (2006) have noted, ‘‘this conclusion remains controversial, as the ELAN has to date been reliably elicited under only a narrow set of conditions involving word category violations of this type’’ (p. 691). The experimental stimuli in the present study did not involve any violation of word category (or of anything else); nevertheless, we interpret the observed early anterior negativity as an ELAN-like response that indexes a key aspect of phrase structure building, and if this interpretation is justified, then it also serves as evidence that the ELAN is not in fact limited to specific violations, but is also elicited in conditions in which legitimate phrase structure is built, perhaps rendering its status somewhat less controversial. Here are the reasons that lead us to suggest that the anterior negativity in response to our conditions is indeed due to an aspect of phrase structure building. Because in the Coercion and Neutral conditions the commitment to C1 as the head-noun has to be revised at C2, there ought to be an associated ERP effect. When the parser encounters C2, it experiences difficulty building the compound-noun phrase because it did not expect the category of this next word to be another noun, having already settled on C1 as the head-noun. C2 constitutes an unexpected word category, but not a word-category violation. In all of the cases in the literature in which an ELAN has been elicited, there has been a strong expectation about what the category of the following word is going to be. We would add that, in certain circumstances, it may also be elicited where there is a strong prediction regarding what category a following word will not be. In *Max’s of proof the theorem, as Lau, Stroud, Plesch, and Phillips (2006) note, Max’s clearly indicates that a noun-phrase is being built and that the head should follow. In our stimuli, the head has already been built, so the parser is startled by the appearance of another noun. Neither Friederici’s (1995) account of ELAN nor Lau et al.’s (2006) alternative to that account envisages an ELAN for wellformed structures. Lau et al. provide a pertinent example: No ELAN occurs for the comparison between (17a) and (17b): (17)

a. b.

The boy embraced Anne’s really nervous cousin The boy embraced Anne really nervously

Note that in (17a), really is an allowable modifier of a typical nominal modifier (i.e., an adjective), that is, it is more or less a predictable element in the extended projection of a noun phrase. In the standard environments that give rise to an ELAN, a word is encountered that is not compatible with the projection of an expected head noun; it has zero predictability. And, with respect to our stimuli, C2 is not compatible with an expected head because that head has already been committed to at C1. A generalization across all of these cases is that the parser notices that an incoming word either poses a problem with respect to the head of a phrase or it does not. On this view, an ELAN is elicited whenever an incoming word is head-incompatible. This tentatively-advanced head-incompatibility view of ELAN is arguably consistent with Lau et al.’s highly plausible suggestion that during such an early time window (under 300 ms), ‘‘only situations where category predictions make diagnosis of illicit incoming words particularly straightforward’’ (Lau et al., 2006, p. 82) would be capable of being noticed by the parser. It is consistent in the sense that the ‘‘space of alternatives’’ (Lau et al., 2006, p. 85) is not appreciably broadened to include syntactic contexts that would be harder to diagnose within 300 ms, though what


counts as ‘straightforward’ is an empirical question. Our findings, however, would suggest a modification of the general current understanding that ELAN occurs only for violations. But detecting a head-incompatibility is all that the ELAN indexes (if that is indeed what it does). It cannot ‘‘determine if there is any way to incorporate the incoming word into the existing parse tree’’ (Lau et al., 2006, p. 85). That ‘repair’ or ‘reanalysis’ ought to occur at a later time, and is the function that we attributed to a P600 that we predicted would be associated with the shift in head-noun from C1 to C2. However, this prediction was not borne out – there was no P600. What we do see, however, is an extended negativity from 350 to 500 ms in AC and AR. This time window does not reveal any sharp distinction from the earlier 150–350 ms window; however, it occurs only in AC and AR, and not in AL, suggesting a distinct process than the earlier 150–350 ms process that was common to all three anterior ROIs. There is a further reason to believe that the 350–500 ms negativity is distinct from the earlier negativity. A reviewer suggested checking the C2 waveforms following a 100 ms baseline (instead of our 200 ms baseline); the findings were the same in all relevant respects, but one subtle difference emerged that is of interest; that is, there was a hiatus between the earlier negativity (significant at 150–350 ms) and the later negativity (significant only at 375– 500 ms).4 It is as if the same neural regions are activated and later reactivated, reinforcing the likelihood of distinct processes being engaged. So, there are two questions to address: first, what might the distinct follow-up process be? Second, why was no P600 observed? A number of studies have reported an ELAN that was not followed up by a predicted P600, and there seems to be no entirely satisfactory account of this phenomenon. Ye, Luo, Friederici, & Zhou, 2006, for example, obtain a sustained negativity in response to phrase-structure violations, but not the predicted P600; it was expected that there should be repair, typically associated with the P600, yet there was no P600. Steinhauer & Drury’s (2012) probing review of ELAN findings raises the possibility that it is probably sustained negativities rather than short-lived ELAN effects that need to be explained; they foster the idea that working memory might play a role. WM effects have often been implicated in language processing that elicits sustained anterior negativities, which can be bilateral in certain contrasts (e.g., King & Kutas, 1995; Phillips, Kazanina, & Abada, 2005), so it is possible that the sustained anterior negativity we observe in the present study is due to an increased burden on WM resources. Garden-pathing has been linked to limitations in WM capacity (King & Kutas, 1995), so it may be that since subjects were garden-pathed into expecting a single-noun, encountering a second noun might engage WM resources to hold C1 in memory while a C1–C2 compound is constructed. In addition, the LAN is known to be constrained by WM capacity (Vos et al., 2001). Based on these considerations, what we might be seeing is the predicted, short-duration ELAN-like response at all three anterior ROIs during the early 150–350 ms time window, merging into a LAN/RAN-like WM-induced negativity restricted to two anterior ROIs during the subsequent 350– 500 ms period. While this is possible, it cannot be confirmed as the study was not designed to test WM effects. Regarding the absence of a P600, it may be that since no actual phrase-structure violation occurred, there was not a need for repair in quite the same way as there would be for actual violations. Also,

4 During the 100–350 ms window, at AC, Coercion and Neutral were significantly more negative than Extra Syntax (p = 0.04 and 0.02, respectively); the same pattern obtained at AR (p = 0.03 and 0.02, respectively). During the 375–500 ms window, again the same pattern: at AC (p = 0.03 and 0.086, respectively); at AR (p = 0.048 and 0.03, respectively).

55

compared to the rather striking phrase-structure violations that have traditionally given rise to an ELAN, one might expect that demands imposed on the parser in the present case would be relatively cost-free, that is, recovery from the structural garden-path in the present study would be quite straightforward, as N–N compounds are so abundant in Icelandic. However, all that can be maintained with confidence is that there was an effect (ELAN-like followed by a more sustained negativity at two ROIs) for the conditions that had to revise head-noun status on encountering C2. The only condition that did not have to revise head-noun status – Extra Syntax – showed no effect at all. Returning now to the idea of no further syntactic work being necessary for the Extra Syntax condition, we can find no study that has investigated the effects of revising head-noun status independently of semantic plausibility or of meaning shifts. Interestingly, though, even in compound studies where plausibility is manipulated, the idea that no further syntactic work has to be done echoes what we postulate for our Extra Syntax condition at C2. For example, Koester, Holle, and Gunter (2009) present subjects with threenoun compounds; that is all they see, just the compounds, no prior sentence context, and the fillers, too, are all three-noun compounds. They manipulate plausibility at C2 and C3. Among other findings, they report a P600 at C2 for less plausible nouns, but none at C3 for either plausible or implausible nouns. They regard the effect at nonhead C2 as a measure of the difficulty of restructuring an implausible noun as a modifier; and they view the absence of a P600 at any C3 as a reflection of the fact that no further restructuring has to be done. In their words, ‘‘As the third constituents were the last constituents of our stimuli, no further adaptation of the compound structure was necessary and no positivity would be expected’’ (2009, p. 1860). The very fact that Koester et al. (2009) obtained an effect for revising the compound structure is intriguing; the fact that they interpret the lack of an effect at C3, when subjects know that they have reached the end, is quite arresting. Making all due allowances for the important differences between their study and ours, there is a useful parallel to be drawn: both studies report head-noun revision effects where they are expected, and none where they are not. Thus, if we place their study and ours side by side, we get a binocular view of the electrophysiological effects of head-noun revision.5

5. Conclusions While the present experiment did not find an ERP effect for mass-count coercion, contrary to predictions, it did show that the effects of building an unexpressed container noun into the syntax are substantial. At the nonhead constituent (C1), the Extra Syntax condition elicited an anterior negativity in the 350–450 ms time window, followed by a P600. We interpret the anterior negativity as the detection of a gender mismatch, and the P600 as a reflection of incorporating the unexpressed container-noun into the syntactic representation. We argue that these findings at C1 dovetail quite neatly with the complementary findings in German compounds. At C2, the Coercion and Neutral conditions exhibit an ELAN in 5 Before leaving the general discussion, we would like to draw attention once again to a phenomenon that we raised in the introduction, namely, the possibility of a quite unusual ambiguity (impossible in English, German, and other languages that we know of) in the Extra Syntax compound. We mentioned that rum bottle could have both bottle of rum and bottle for rum interpretations, but that we had provided pragmatic contexts that permitted only the bottle of rum reading. Now, suppose that instead of controlling the ambiguity as we did, we manipulate it, which can be done by choice of verb, then we can contrast ‘drink a coffee cup’ and ‘break a coffee cup’, where the verb predicates itself of either C1 or C2, respectively. Future work will investigate the intriguing possibilities that this contrast opens up for processing accounts involving headedness (semantic as well as syntactic) and position-in-string.

56


comparison with the Extra Syntax condition. We make a probationary case that the ELAN is an index of the apparent head-incompatibility of C2; and we conjecture that this is not seen for the Extra Syntax condition because its syntactic processing is already complete, and in this connection we draw attention to another parallel in research on German compounds. Acknowledgments We are indebted to many people without whose contributions this study would have been impossible; although we cannot acknowledge all of them here, we should at least mention the following: We would like to thank the 23 subjects at the University of Iceland who took part in the experiment; Þórhallur Eyþórsson and his students, Brynhildur Stefánsdóttir and Sigríður Sæunn Sigurðardóttir, for helping in stimuli construction; Þór Eysteinsson and Einar Jón Einarsson for offering us full, unfettered use of their laboratory facilities; Jóhannes Gísli Jónsson and Þórhallur Eyþórsson for assistance with administering pretests; Sigrún Helgadóttir and Steinþór Steingrímsson for their corpus work on frequency counts; to all of the brilliantly engaged students at the MSU EEG Lab meetings who made so many insightful suggestions over many months; to Dirk Koester and Heike Wiese for a thorough critical reading of an earlier version of the paper; to Jason Moser and Colin Phillips for consistently useful consultation and advice at different stages of the project; and to Karin Wurst, David Prestel, and Tom Lovik for their material assistance. To all of these people, we are sincerely appreciative. None of them is responsible for the errors that remain. The material is based in part on work done while the fifth author was serving as Director of the Linguistics Program at the U.S. National Science Foundation. Any opinions, findings, and conclusions expressed in this material are those of the authors, and do not necessarily reflect the views of the U.S. National Science Foundation. References Allen, M., Badecker, W., & Osterhout, L. (2003). Morphological analysis in sentence processing: An ERP study. Language and Cognitive Processes, 18, 405–430. Baggio, G., Choma, T., van Lambalgen, M., & Hagoort, P. (2009). Coercion and compositionality. Journal of Cognitive Neuroscience, 22, 2131–2140. Barber, H., Salillas, E., & Carreiras, M. (2004). Gender or genders agreement? In M. Carreiras & C. Clifton (Eds.), On-line study of sentence comprehension; eyetracking, ERP and beyond. Brighton, U.K.: Psychology Press. Borer, H. (2005). In name only. Structuring sense (Vol. 1). Oxford: Oxford University Press. Copestake, A., & Briscoe, R. (1995). Semi-productive polysemy and sense extension. Journal of Semantics, 12, 15–67. Coulson, S., King, J., & Kutas, M. (1998). Expect the unexpected: Event-related brain response to morphosyntactic violation. Language and Cognitive Processes, 13, 21–58. Daðason, J. F. (2012). Post-correction of Icelandic OCR text. M.Sc. Thesis, Faculty of Industrial Engineering, Mechanical Engineering and Computer Science, University of Iceland. Delogu, F., Crocker, M., & Drenhaus, H. (2012). When coercion comes as a surprise: An ERP study in German. Poster presented at NLC, San Sebastian, Spain (October 25–27). Deutsch, A., & Bentin, S. (2001). Syntactic and semantic factors in processing gender agreement in Hebrew: Evidence from ERPs and eye movements. Journal of Memory and Language, 45, 200–224. Dillon, B., Nevins, A., Austin, A. C., & Phillips, C. (2012). Syntactic and semantic predictors of tense in Hindi: An ERP investigation. Language and Cognitive Processes, 27, 313–344. Friederici, A. D. (1995). The time course of syntactic activation during language processing: A model based on neuropsychological and neurophysiological data. Brain and Language, 50, 259–281. Friederici, A. D. (2002). Towards a neural basis of auditory sentence processing. Trends in Cognitive Science, 6, 78–84. Friederici, A. D., & Frisch, S. (2000). Verb argument structure processing: The role of verb-specific and argument-specific information. Journal of Memory and Language, 43, 476–507. Frisson, S., & Frazier, L. (2005). Carving up word meaning: Portioning and grinding. Journal of Memory and Language, 53, 277–291.

Gouvea, A. C., Phillips, C., Kazanina, N., & Poeppel, D. (2010). The linguistic processes underlying the P600. Language and Cognitive Processes, 25, 149–188. Gunter, T. C., Friederici, A. D., & Schriefers, H. (2000). Syntactic gender and semantic expectancy: ERPs reveal early autonomy and late interaction. Journal of Cognitive Neuroscience, 12, 556–568. Gunter, T. C., Stowe, L. A., & Mulder, G. (1997). When syntax meets semantics. Psychophysiology, 34, 660–676. Hagoort, P., Brown, C., & Groothusen, J. (1993). The syntactic positive shift as an ERP measure of syntactic processing. Language and Cognitive Processes, 8, 439–483. Hagoort, P., Brown, C., & Osterhout, L. (1999). The neurocognition of syntactic processing. In C. Brown & P. Hagoort (Eds.), The neurocognition of language (pp. 273–316). Oxford: Oxford University Press. Hagoort, P., Wassenaar, M., & Brown, C. (2003). Syntax-related ERP-effects in Dutch. Cognitive Brain Research, 16, 38–50. Hahne, A., & Friederici, A. D. (2002). Differential task effects on semantic and syntactic processes as revealed by ERPs. Cognitive Brain Research, 13, 339–356. Helgadóttir, S., Svavarsdóttir, Á., Rögnvaldsson, E., Bjarnadóttir, K., & Loftsson, H. (2012). The tagged Icelandic corpus (MÍM). In Proceedings of language technology for normalization of less-resourced languages, workshop at the 8th International conference on language resources and evaluation, LREC, 2012, Istanbul, Turkey. Jackendoff, R. (1997). The architecture of the language faculty. Cambridge, MA: MIT Press. King, J. W., & Kutas, M. (1995). Who did what and when? Using word- and clauselevel ERPs to monitor working memory usage in reading. Journal of Cognitive Neuroscience, 7, 376–395. Koester, D., Gunter, T. C., & Wagner, S. (2007). The morphosyntactic decomposition and semantic composition of German compound words investigated by ERPs. Brain and Language, 102, 64–79. Koester, D., Gunter, T., Wagner, S., & Friederici, A. D. (2004). Morphosyntax, prosody, and linking elements: The auditory processing of German nominal compounds. Journal of Cognitive Neuroscience, 16, 1647–1668. Koester, D., Holle, H., & Gunter, T. C. (2009). Electrophysiological evidence for incremental lexical-semantic integration in auditory compound comprehension. Neuropsychologia, 47, 1854–1864. Kuperberg, G. R., Choi, A., Cohn, N., Paczynski, M., & Jackendoff, R. (2010). Electrophysiological correlates of complement coercion. Journal of Cognitive Neuroscience, 22, 2685–2701. Kutas, M., Van Petten, C. K., & Kluender, R. (2006). Psycholinguistics electrified II (1994–2006). In M. A. Gernsbacher & M. Traxler (Eds.), Handbook of psycholinguistics (2nd ed., pp. 659–724). New York: Elsevier Press. Lau, E., Stroud, C., Plesch, S., & Phillips, C. (2006). The role of structural prediction in rapid syntactic analysis. Brain and Language, 98, 74–88. Molinaro, M., Vespignani, F., & Job, R. (2008). A deeper reanalysis of a superficial feature: An ERP study on agreement violations. Brain Research, 1228, 161–176. Montague, R. (1970). Universal grammar. Theoria, 36, 373–398. Münte, T. F., & Heinze, H. J. (1994). ERP negativities during syntactic processing of written words. In H. J. Heinze, T. C. Münte, & G. R. Mangun (Eds.), Cognitive electrophysiology (pp. 211–238). Boston: Birkäuser. Neville, H. J., Nicol, J. L., Barss, A., Forster, K. I., & Garrett, M. F. (1991). Syntacticallybased sentence processing classes: Evidence from event-related potentials. Journal of Cognitive Neuroscience, 3, 151–165. Osterhout, L., & Holcomb, P. J. (1992). Event-related brain potentials elicited by syntactic anomaly. Journal of Memory and Language, 31, 785–806. Osterhout, L., & Mobley, L. A. (1995). Event-related potentials elicited by failure to agree. Journal of Memory and Language, 34, 739–773. Pakulak, E., & Neville, H. J. (2010). Proficiency differences in syntactic processing of monolingual native speakers indexed by event-related potentials. Journal of Cognitive Neuroscience, 22, 2728–2744. Pelletier, F. J. (1975). Non-single reference: Some preliminaries. Philosophia, 5, 451–465. Pelletier, F. J. (2010). Mass terms: A philosophical introduction. In F. J. Pelletier (Ed.), Kinds, things, and stuff: Mass terms and generics (pp. 123–131). Oxford: Oxford University Press. Phillips, C., Kazanina, N., & Abada, S. H. (2005). ERP effects of the processing of syntactic long-distance dependencies. Cognitive Brain Research, 22, 407–428. Piñango, M. M., Zurif, E., & Jackendoff, R. (1999). Real-time processing implications of enriched composition at the syntax–semantics interface. Journal of Psycholinguistic Research, 28, 395–414. R Core Team. (2013). R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing, http://www.R-project.org. Ramchand, G. (2008). Verb meaning and the lexicon: A first phase syntax. Cambridge: Cambridge University Press. Roehm, D., & Haider, H. (2009). Small is beautiful: The processing of the left periphery in German. Lingua, 119, 1501–1522. Rösler, F., Pütz, P., Friederici, A. D., & Hahne, A. (1993). Event-related brain potentials while encountering semantic and syntactic constraint violations. Journal of Cognitive Neuroscience, 5, 345–362. Staub, A., Rayner, K., Pollatsek, A., Hyönä, J., & Majewski, H. (2007). The time course of plausibility effects on eye movements in reading: Evidence from noun–noun compounds. Journal of Experimental Psychology: Learning, Memory, and Cognition, 33, 1162–1169. Steinhauer, K., & Drury, J. E. (2012). On the early left-anterior negativity (ELAN) in syntax studies. Brain and Language, 120, 135–162. Ueno, M., & Kluender, R. (2009). On the processing of Japanese wh-questions: An ERP study. Brain Research, 1290, 63–90.

M. Whelpton et al. / Brain & Language 136 (2014) 44–57 Vos, S. H., Gunter, T. C., Kolk, H. H., & Mulder, G. (2001). Working memory constraints on syntactic processing: An electrophysiological investigation. Psychophysiology, 38, 41–63. Weyerts, H., Penke, M., Dohrn, U., Clahsen, H., & Münte, T. (1997). Brain potentials indicate differences between regular and irregular German plurals. Neuroreport, 8, 957–962. Wiese, H., & Maling, J. (2005). Biers, Kaffi, and Schnaps: Different grammatical options for restaurant talk coercions in three Germanic languages. Journal of Germanic Linguistics, 17, 1–38.

57

Wood, J. (2012). Icelandic morphosyntax and argument structure. Ph.D. dissertation, Department of Linguistics, New York University. Yamada, Y., & Neville, H. J. (2007). An ERP study of syntactic processing in English and nonsense sentences. Brain Research, 1130, 167–180. Ye, Z., Luo, Y., Friederici, A. D., & Zhou, X. (2006). Semantic and syntactic processing in Chinese sentence comprehension: Evidence from event-related potentials. Brain Research, 1071, 186–196.

Disentangling Metaphor from Context: An ERP Study.

An ERP-study of brand and no-name products.

Social distance and anonymity modulate fairness consideration: An ERP study.

Processing Metrical Information in Silent Reading: An ERP Study.

Inferential revision in narrative texts: An ERP study.

verb distinction in English stress homographs: an ERP study.

Anticipatory processes under academic stress: an ERP study.

The neural response to maternal stimuli: an ERP study.

Rendaku (Japanese sequential voicing) as rule application: an ERP study.

Neural manifestations of implicit self-esteem: an ERP study.

An ERP study on initial second language vocabulary learning.

Knowledge of sequence structure prevents auditory distraction: an ERP study.

Impact of Semantic Relatedness on Associative Memory: An ERP Study.

On content of practice. An Icelandic multicentre study, population, practices and contacts.

Genetic changes in breast carcinomas in an Icelandic population.

An automated procedure for correlation of q-sorts.

RNA: All sorts of mitochondrial RNA.

Gender differences in reward and punishment for monetary and social feedback in children: An ERP study.

An evidence for lack of pseudoneglect in patients with schizophrenia: an ERP study.

Beliefs in being unlucky and deficits in executive functioning: an ERP study.

Reasons for contact in family practice. An Icelandic multicentre study on content of practice.

The Temporal Courses of Phonological and Orthographic Encoding in Handwritten Production in Chinese: An ERP Study.

Infants' object location and identity processing in spatial scenes: an ERP study.

Audio-visuomotor processing in the musician's brain: an ERP study on professional violinists and clarinetists.