That nouns and verbs are different is perhaps the most commonly-known metalinguistic fact. "Verb: that's what's happening!" (Dorough and Yohe, 1974) and "A noun is a person, place, or thing." (Ahrens and Sidebotham, 1973) are "naive grammar" familiar to every schoolchild; the distinction is maintained throughout even the most remote arcana of modern linguistics. All languages have both types of words, and employ them in some similar ways; this appears to be a universal feature of human languages. Both nouns and verbs can be described as "open-class" words -- that is, every language contains quite a large number of each, and people often continue to learn new ones throughout their lives.

For example, the Projection Principle of formal linguistic theory distinguishes between the roles nouns and verbs play in sentence construction. (For a good introduction, see Haegeman (1991).) A verb is considered to set up (or "project") some number of thematic roles, which must then be filled by an appropriate number of nouns. The verb "smile" projects one role -- that of the agent, who smiles, so "Susan smiles." is a grammatically complete sentence. "Susan smiles the puppy." is not a good sentence; if we want to get another noun in there, we need to also add some intervening construction ("Susan smiles at the puppy.") in order to achieve syntactic acceptability. The verb "likes", on the other hand, projects two roles -- that of the one doing the liking, and that of the liked; "Susan likes." is not much of a sentence. "Susan likes the puppy." is appropriately formed. So is another sentence -- "The puppy likes Susan." -- but here the roles played by the nouns have been reversed. Note that the noun words are the same in all these examples; however, the ways in which the nouns can or cannot appear are dictated by the particular verb.

Within a given language, nouns and verbs can also sometimes be differentiated by cues that are not necessarily specifically syntactic or semantic, such as noun- or verb-marking phonology. For example, there is a certain class of English words in which the stress pattern indicates noun or verb -- a thing that people "reJECT" (verb) is then known as a "REject" (noun). Kelly (1992) documents the extent of these, and reviews experimental evidence that speakers are aware of such distinctions.

Nouns and verbs can also be considered as two different types of cognitive stimuli, and a number of experiments have investigated the ways in which they are organized, processed, and remembered. For example, Engelkamp, Zimmer, and Mohr (1990) demonstrated that lists of concrete nouns were easier to remember and organize than similar lists of action verbs. Mohr-Gilbert (1992) demonstrated that in a paired-words memory task, nouns cue verbs better than verbs cue nouns. Moreover, an experimental condition in which subjects acted out the thing that the word-pair described (as opposed to merely memorizing the pair in the abstract), verb-cue performance showed more improvement than noun-cue performance did. Findings such as these demonstrate that the syntactic category of a word affects cognition even outside the syntactic context of a sentence.

As one progresses from infancy to linguistic competence, one must learn a great many things about one's languages, including nouns, verbs, and various rules for dealing with each. Although the typical adult will know both nouns and verbs and handle them with competence, early language often includes differences between noun and verb development.

Several factors seem to contribute to these differences, including the nature of the language being learned, and the particular motherese to which the child is exposed. Children learning a language in which nouns are paramount, and whose caregivers use a lot of nouns in speaking to the child, generally begin using nouns earlier than verbs. That this is generally the case in English has been well-documented, and many theories developed as to why this should be so (e.g. Gentner, 1978). However, other languages with different features -- such as Korean and Japanese (Gopnik and Choi, 1995) -- or children whose caregivers have a different style of speaking to the child (Goldfield, 1990), may show different patterns during development.

Nonetheless, the distinction between syntactic classes appears to be important during the development of any language. Maratsos (1988) points out that merely semantic distinctions are not sufficient to explain children's development of syntactic categories; for example, not all verbs are going to be action-related, while many adjectives and prepositions are. Children typically apply the verb rule of adding -ed to signify past tense even to non-action verbs such as "know" or "want", but not to even action-related adjectives like "helpful".

Carey and Bartlett (1978) demonstrated the capability of young children to learn a new word from as little as a single exposure to it in a real-world context. They presented a color word ("chromium", used to designate olive green) to twenty three-year-olds, in the context of a participatory discrimination task ("Bring me the chromium tray -- not the red one.") As many as 65% of their subjects subsequently demonstrated at least some learning of the word on a comprehension task. Merriman and Stevenson (1997) have shown a "small but reliable" tendency to show some learning of noun classes from a single novel exposure in children as young as 24-25 months.

Children also appear to use quite a bit of the surrounding syntactic information surrounding to learn about unfamiliar words. (This is called "syntactic bootstrapping".) Naigles (1990) presented two-year-olds with a pair of different kinds of actions and a nonsense word in one of two different (verb) syntactic structures; when the children were later asked to choose which action the verb went with, their choice was a function of the syntactic structure they had heard.

In general, open-class word learning can take place at any time in life, and little effort is associated with it -- learning the "jargon" is a major part of taking up any new hobby or new area of interest. A single exposure is often all that's necessary.

Pseudoword stimuli, presented in syntactically and semantically rich contexts, place some interesting demands on the language processor. They are not, of themselves, either nouns or verbs -- neither syntactic nor semantic information about them can be obtained from the lexicon. But their semantics and syntax can be inferred, to some extent, from the context in which they are presented -- in "She took a sip of the spoo." the pseudoword is clearly a noun, moreover a liquid, and most likely a potable one; in "The children spoo all day.", the same stimulus must function as a verb. This is syntactic bootstrapping all over again, rather as it was for Naigles' (1990) two-year-olds.

Spenney and Haynes (1989) trained English-speaking university students in novel "object words" and "action words", and tested reaction times for retrieval. Object words (nouns) came easier to these subjects, both for comprehension and production. However, no comparable study has been done in Korean, Japanese, or another verb-facilitative language.

A number of studies have documented differences between the relative impairment of nouns and verbs in brain-damaged individuals. Miceli et. al. (1984) examined verb omission in a group of Italian speakers with brain damage affecting their language use. "Agrammatic aphasics", whose primary symptomology was characterized by the omission of syntactic markers and reduced phrase length relative to matched controls, were compared with other patients described as "anomics", whose syntactic production was fluent but whose speech was characterized by anomias, circumlocutions, and obvious word-finding difficulty in their spontaneous speech. A linguistically-intact control group participated as well. As a group, agrammatics performed better on noun tasks than on verb ones, while anomics performed better on verbs than on nouns. Interestingly, normal controls also performed better on verbs. (It is perhaps also worth noting that in Italian, the overwhelming majority of words are distinctively and solely either nouns or verbs; few words in Italian have both noun and verb usages, whereas many English words do.)

Zingeser and Berndt (1990) examined the relative impairment of nouns and verbs as related to the broader aphasic categorizations of agrammatism and anomia. Nouns and verbs were elicited as single words in picture-naming, naming-to-definition, and (for verbs) action description tasks. Narrative speech was also studied with a story elicitation. It was found that agrammatics produced fewer verbs than nouns in all the tasks, while anomics produced more verbs than nouns in naming-to-definition. This replicated a similar, earlier finding in Italian. The pattern seems to hold even outside the Indo-European language family. Bates et. al. (1992) studied Broca's and Wernicke's aphasias in speakers of Chinese -- a language that has neither declensions nor conjugations of its verbs. This removes the correlation of morphological load and verbness that is present in, for example, English. Again, action-naming -- a verb function -- was deficient in the Broca's patients, and relatively spared for the Wernicke's aphasics.

Caramazza and Hillis (1991) report two very interesting cases of verb deficit -- one with principally verb-related deficits in oral production, the other with problems virtually only in verbs in written production. This double dissociation -- documented in read-aloud and write-to-dictation conditions, as well as a picture description task -- implies that physiologically separate lexicons for verbs exist in the phonological and orthographic output lexicons. The oral-deficit patient also showed semantic paraphasias (i.e., reading "dollar" as "money"); the written-deficit patient showed some morphological errors in complex words (i.e., reading "darken" as "darkness"). Unfortunately, data were not provided to help determine to whether these other deficits were related at a processing level, or whether they merely reflected unassociated damage to the subjects' brains.

Damasio and Tranel (1993) described two patients who could produce verbs at normal levels, while their noun production was extremely impaired. A third patient performed well on the nouns task, but verb retrieval was defective. The noun-impaired patients had both sustained damage to left anterior and middle temporal lobe (and each, independently, to other areas of cortex as well) -- that is, outside the so-called "language areas" commonly associated with aphasias. Indeed, their linguistic performance was otherwise good -- "there were no impairments in grammar, morphology, phonetic implementation, or prosody; reading and writing were normal". This may point up again the different roles that nouns and verbs play -- brain damage can apparently impair nouns without impacting sentence construction, but the same may not be true for verbs.

Daniele et. al. (1993, 1994) noted three patients whose deficits were specific to one or another syntactic category: one with left temporal lobe atrophy who had trouble with both naming and comprehending nouns; two with left frontal damage (one perfusion abnormality; one atrophy) who had difficulty naming and comprehending verbs.

A striking case of syntactic-category-specific impairment was reported by Hillis and Caramazza (1995), who described a patient whose left frontal and tempoparietal strokes impaired her picture-naming far more for nouns than for verbs, but whose ability to read written verbs was severely impaired, while her noun-reading remained quite intact. This set of dissociations -- between nouns and verbs, and between written input and spoken output -- in a single patient is most striking evidence of underlying physiological distinctions between such processes. They cannot be due simply to differences of "difficulty" between one syntactic category and another, since performance on each is relatively preserved in different types of task.

Warburton et. al. (1996) used positron emission tomography to identify significant local changes in regional cerebral blood flow (rCBF) during a number of tasks in which the subjects were instructed to think of nouns and verbs. These were "word-level" tasks, where the subjects were told to generate lists of words of one kind or another (rather than full sentences); they were also told, for each task, whether to generate nouns or verbs (so the tasks depended also on the subjects' metalinguistic knowledge). Some of the tasks involved word repetition; others involved semantically-based word generation (e.g., in a noun task, a subject might be asked to generate "basic-level" nouns from a category -- so given "furniture", to think, "chair, table, cabinet"; in a verb task, given "apple", to generate related verbs such as "eat" and "peel"). (Note that both of those tasks began with the implicit task of comprehending a noun.)

In general, these studies found extensive activation of the left prefrontal cortex, and medially, the anterior cingulate cortex and supplementary motor area (SMA). When compared with rest states, retrieval of nouns and verbs were associated with the same distribution of activation. When compared to each other, verb generation produced greater activation in posterior and anterior regions of the lateral left hemisphere, and there was a trend for greater activation in the SMA. Noun generation was associated with only a few regions of greater activation, one in the right prefrontal cortex -- at a locus where verb generation was associated with a decrease in rCBF as compared with rest.

Also of potential relevance are a few studies (Martin et. al. 1996, Damasio et. al. 1996), which examined rCBF in naming tasks involving various semantically-defined noun subcategories (people, animals, tools) in normal and brain-damaged subjects. All the naming tasks apparently involved both "classical language areas" such as Broca's area in the left hemisphere; additionally, various categories apparently needed to recruit particular other areas -- naming animals selectively activated the left medial occiptial lobe, which is involved in the earliest stages of visual processing; tools, by contrast, selectively activated a left premotor area also activated by imagined hand movements.

While there are certainly many nouns that are not particularly visual ("noise") and verbs that are not particularly motor ("remember"), it is nevertheless possible that there are pervasive (if not entirely systematic) semantic as well as syntactic distinctions to be found between the noun and verb lexicons of a language. Since such semantic variations are physiologically distinguishable even within a single semantic category, care should be taken to avoid confounding syntactic category with semantic features.

A basic axiom of cognitive neuropsychology is that brains and thoughts and behaviors are all linked: that changes in one will be reflected by changes in the others. Electropsychological research into cognition is grounded in this notion; it examines a physiological measurement of EEG activity in the context of thought- or behavior-based events. This approach enables psycholinguists to investigate what one's brain is doing while one is doing language.

ERPs (Event-Related brain Potentials) are an average of EEG signals time-locked to the onset of some set of sensory, motor, or cognitive events (Hillyard & Picton, 1987). The averaged electrical waveform, plotted as change in charge over time, consists of a series of voltage peaks called the ERP "components". ERPs have been shown to be consistently sensitive to a variety of stimulus manipulations; there are several components that are sensitive to manipulations of linguistic features (Osterhout, 1994). ERPs have several valuable features as research tools; they are an on-line, physical measurement that gives a multidimensional, millisecond-by-millisecond reflection of brain processes, and they do not require the subject to perform some additional behavioral task which might potentially interfere with or contaminate the cognitive process of interest.

The appropriate interpretation of ERP responses must take into account their correlative nature, and not impute simple causality to the expermental manipulation. If one notes that, say, syntactic gender disagreement evokes a relatively postive-going wave 600 milliseconds post-stimulus-onset, that does not necessarily mean that the P600 has a one-to-one relationship with gender disagreement. Other kinds of stimuli (say, number disagreement) may evoke quite the same effects.

In terms of experimental manipulations that do have an effect, ERP responses may be divided into "exogenous" and "endogenous" types -- that is, those that vary according to physical manipulations of stimulus features such as size or color, and those that change to reflect cognitive variables such as expectation. Since "language" is a phenomenon of the mind, its essential processes will be reflected within the endogenous componentry. But since all language is conveyed through physical media (such as sounds, or physical gestures, or written symbols), exogenous components can be affected by non-linguistic variables in the stimuli as well! In order to distinguish between the two, it will be necessary to design stimuli carefully, so that the physical and linguistic aspects of the stimuli do not vary together in any systematic way.

A note on the names of components: By convention, ERP components are referred to by names containing a letter and a number. The letter, N or P, designates whether the wave is negative- or positive-going in polarity. There are two different conventions for the number -- one referring to the order in which the components appear (so a "P3" is the third positive-going wave), and the other referring to the number of milliseconds between stimulus onset and component's occurrence (so the "N400" is a negative-going wave with a peak 400 milliseconds after stimulus presentation). I have generally used the latter convention in this paper, though it is somewhat complicated by (a) changes in component timing between experiments, related to factors such as changes in stimulus presentation rates and (b) some inconsistency in referent, whether to a wave's peak or the time at which a difference first becomes apparent (some waves, for example the "P600", often don't even have clearly defined peaks).

A note on the ERP diagrams: All ERP waveforms shown here have negative voltage plotted upwards. Unless otherwise noted on the plot, the vertical calibration bar represents 5 microvolts. Time is plotted on the horizontal axis; each tic represents 100 milliseconds, and the onset of presentation of critical words is at the vertical calibration bar.

The ERP associated with comprehending a sentence is essentially sinusoidal in shape. Each word generates one period of the curve, a series of distinctive positive- and negative-going brainwaves (the "components") that remains similar no matter what the language is, or whether the material presented is written, spoken, or signed. Certain other things about the words, though, can have enormous effects upon particular parts or aspects of the waveform, such as the peak size or latency of a particular component. Sentence position, grammatical class, the subject's linguistic background, and various types of syntactic and semantic anomaly have all been shown to have systematic effects upon various parts of this basic wordshaped brainwave. Hypotheses about language processing can be evaluated in the light of these physiological variations. Such evidence can form the basis for a physiologically-motivated taxonomy of the cognition of language comprehension.

Linguistic aspects of a stimulus have been associated with endogenous components in the ERP. The linguisticness of a stimulus reflects purely cognitive variables -- non-signers, for example, might not even recognize a conversation in American Sign Language as a linguistic activity! But many linguistic features within a comprehended language can affect the electrical activity measured by the ERP.

The earliest component that seems to be responsive to linguistic variables is the P2 (also called the P200). As the names imply, this is a positive-going wave; it is the second one generally observed in response to stimulus presentation, and it peaks close to 200 milliseconds post stimulus presentation. While it is apparently responsive to some non-linguistic variables, such as physical complexity and relevance of pictures (Carretie and Iglesias (1995)), it has also shown some responsiveness to syntax.

Experiments by Neville et. al. (1991) included a very wide variety of syntactic manipulations, some of which significantly affected P2 amplitude. Response to the word "criticized" in sentences such as "What was a proof of criticized by the scientist?" showed a larger P2 than response to the same word in "Was the proof of the theorem criticized by the scientist?" (a syntactic-context difference characterized by the authors as a "subjacency constraint violation"). A similar effect was seen on the response to "of" in the anomalous "The scientist criticized Max's of proof the theorem." relative to "The scientist criticized a proof of the theorem." (Neville et. al.'s "phrase structure violation" case).

McKinnon and Osterhout (1996) showed that the word "when" evoked greater P2 positivity when it appeared in a sentence that violated "subjacency constraints" (such as "I wonder which of his staff members the candidate was annoyed when his son was questioned by.") than in controls (such as "I wonder whether the candidate was annoyed when his son was questioned by his staff members."). In these experiments, P2 peak amplitude was not only greater, but unlike the responses seen in Neville et. al. (1991), the positivity persisted throughout the rest of the recorded epoch.

If the stimulus is an open-class word, then after the P2, the ERP begins a steep climb towards the negative. This wave, which starts around 250 milliseconds and reaches a peak about 400 msec after stimulus onset, is called the N400. It is generally largest at central-parietal scalp locations. It was the first component identified as responsive to a linguistic feature (Kutas & Hillyard, 1980), and remains the most studied linguistic ERP component to date.

Kutas & Hillyard (1980) noted a larger N400 response to contextually inappropriate words in sentence-final position: "he spread the warm bread with socks." The size of the N400 was proportional to the degree of semantic incongruity, so "He took a sip from the glass." led to a small N400, "...from the waterfall." a larger one, and "....from the transmitter" largest of all. In Kutas & Hillyard (1984) the relationship was further defined: the size of the N400 was inversely proportional to the Cloze probability of the word in its context. (Cloze probability is measured by presenting a group of subjects with the context as a sentence completion task: "He took a sip from the ________.", and seeing what percentage of the subjects fill in the blank with the particular word. In this case, "glass" and "cup" would have a high Cloze probabilities, "waterfall" a low one, and "transmitter" presumably one around zero.)

Certain other manipulations refined this basic point of knowledge: the specific contextual semantics of the words were involved. Van Petten & Kutas (1987) compared N400s for primed and unprimed semantic relations: "The gambler pulled an ace from the bottom of the deck." followed by "cards" or "ship". Words like "ship" -- related only polysemously to the last word, not to the overall semantic content of the sentence -- yielded larger N400s than those which related to the sentence's general meaning, indicating that semantic contexts produced by the whole sentence (rather than by the single last preceding word) affected N400 amplitude. Van Petten & Kutas (1990) showed that N400 amplitude generally reduced from word to word over a sentence, presumably as semantic context, constraints, and predictability built up.

Chwilla et. al. (1995) showed that the N400 is sensitive to task-related demands as well. In a lexical-decision task, pseudoword stimuli evoked increases in the N400. In another task in which subjects were asked to discriminate capitalized from lower-case words, N400 was not affected.

An N400-like effect at the sentential level has been noted by Osterhout & Holcomb (1992), and replicated in Osterhout & Mobley (1995) and McKinnon & Osterhout (1996), etc. They compared grammatically acceptable and unacceptable sentences ("The broker hoped to sell the stock." versus "The broker persuaded to sell the stock.") and found an increased N400 to the final word ("stock." )of the ungrammatical sentences. Even when the sentences were extended further -- as in "The broker persuaded to sell the stock was sent to jail." -- it was always the final word (marked in written English with punctuation such as a period, or in spoken English by prosodic cues) that evoked the N400-like effect -- so in this longer sentence, the effect would be observed not on "stock" but rather on "jail.". (See next section to find out more about what else occured in such sentences.) This indicates a way that the ends of sentences are cognitively special.

Garnsey, Tanenhaus, and Chapman (1989) demonstrated that increased negativity peaking around 400 milliseconds post-stimulus could also be evoked by stimulus manipulations that might be considered more syntactic than semantic in nature. They measured ERP responses to verbs that could take direct objects, in sentences in which a noun that might belong in the direct-object position preceded the verb (through wh-movement). For example, "The businessman knew which [customer/article] the secretary called....." Here, "customer" is a semantically appropriate filler for the direct object of "called", and "article" is not -- a good role-assignment for "article" can only be made after more context is presented ("....which article the secretary called about.") The N400 response to the verb was larger when it was preceded by the bad direct object ("article") compared to the good one ("customer"). However, it is difficult to say whether this is due to any effect of the parsing process per se, or only that "customer" is more semantically associated with "called" than "article" is in this context. This could be controlled for by using highly-associated but nonetheless direct-object-inappropriate words, as in "The businessman knew which telephone the secretary called on.", but Garnsey et. al. made no measure of this factor in their study.

Of the many syntactic manipulations examined in Neville et. al. (1991), two -- the "phrase structure violation" ("The scientist criticized a proof of ..." versus "The scientist criticized Max's of ...") and the "specificity constraint violation" ("The scientist criticized Max's proof ...." versus "What did the scientist criticize Max's proof ....") -- shared an interesting feature. Although the two sets of waves were basically quite different-looking (as one might expect from a closed-class word set and an open-class one), both evoked significant negativity, largest at left anterior sites, in response to the versions that the experimenters considered violating.

Such effects as these differ from the "classic" N400 not only in the kind of stimuli that generate them, but in their scalp distribution (being primarily left-hemisphere and anterior, rather than largest posterior), perhaps indicating a different underlying neural source, and hence an identifying name: the "Left Anterior Negativity", or LAN.

Munte et. al. (1993) showed subjects German word pairs. In their "syntactic" condition, an article or pronoun was followed by an agreeing or disagreeing noun; in their "semantic" condition, the words in the pair were both open-class, and either semantically related or not. They observed a classical N400 effect in their semantic condition; relative to that, they claim to have observed that their syntactically violating condition evoked a slightly later negativity largest at left frontopolar sites. (However, careful examination of their stimuli, analyses, and figures reveals a somewhat more complex picture. For example, some of their "syntactically anomalous" stimuli were not necessarily so; for example, the construction "you <noun>" may be unacceptable (as in the example they provide, "you parliament"), but is not necessarily so because of syntax (consider the syntactically acceptable, fully-sentential "You animal!"). Furthermore, they report but do not discuss statistically significant results both in epochs earlier than the N400's (during the 100-300 millisecond range, visible in their figures as an increase in P2 amplitude to their "syntactically anomalous nouns" conditions, which included), and later (during the 500-700 millisecond epoch, and apparent in their figures as a positivity in response to syntactic violations in both noun and verb conditions). )

The LAN has also been seen appearing alongside the P600 (about which more later). An example is a report by Osterhout and Holcomb (1992), of a left anterior negativity in the 300-500 millisecond post-stimulus epoch, in response to auxiliary verbs which were "expected to be difficult to attach to the computed sentence structure" -- contrasting cases such as "The banker persuaded to sell the stock was ...." with "The banker hoped to sell the stock was ...". The LAN has also been observed alongside the P600 in such classically syntactic violating cases as failures of number agreement between subject and verb:

Interestingly, other cases of failure to agree in this study (number and gender disagreements between reflexives and their antecedents, as in "The guests helped himself" and "The woman congratulated himself....") evoked only the later positivity, with no evidence of a LAN effect.

So what about that later positivity, then?

Another literature has described language-sensitive ERP phenomena that are positive-going and occur after the N400's peak. These are called, depending upon the investigating laboratory, the Syntactic Positive Shift, or Late Positive-Going Components, or simply the P600.

The P600 is a positive-going wave, reaching its maximum around 600 milliseconds post-stimulus, with an onset around 500 msec. It is apparently sensitive not to the semantic appropriateness or inappropriateness of a given word, but rather to the ease with which the word can be syntactically integrated into the sentence as constructed thus far. It has been demonstrated with a wide variety of stimulus types, including both open- and closed-class words.

For example, Osterhout and Mobley (1995) demonstrated that errors in the bound morphological markers that mark number agreement cause a large, late, positive-going wave:

Note that there is not much increase in N400 amplitude here. Osterhout & Mobley (1995) demonstrated similar effects to gender and number violations in anaphor binding ("The hungry guest helped themselves to the food in the refrigerator.") In terms of gender, the P600 was observed when a definite disagreement took place, as in "The actress looked at himself in the mirror." (It may be particularly worth noting that this effect was documented in English, a language with a quite limited and rudimentary system of gender in its grammar.)

On a more structural syntactic level, Osterhout & Holcomb (reading task, 1992; listening task, 1993) demonstrated that the P600 is sensitive to verb subcategorization and phrase-structure constraints. They used stimuli that were difficult to parse, leading subjects down a "garden path" of assigning syntactic roles, which would need to be reconstructed in order to make sense in light of later context. For example, when a sentence starts with a noun and is followed by a transitive verb, as in "The broker persuaded....", then it is tempting to assume that the broker is the Agent of "persuaded", and that the next thing encountered will be something that will fit into "persuaded"'s object position (as in "The broker persuaded George..."). However, if the sentence goes on: "The broker persuaded TO...." then a reparse is necessary, assigning a reduced relative clausal role: "The broker (who was) persuaded to sell the stock was sent to jail."

For contrast, Osterhout & Holcomb used intransitive verbs: "The broker hoped...." Such verbs are easy to parse in the "to" condition: "The broker hoped to sell the stock." At the word "to", the transitive verbs caused a P600 effect. But if the sentence continued: "The broker [hoped/persuaded] to sell the stock WAS...." then, at the word "was", the intransitive word (which could not be reparsed into a reduced relative) was responsible for increased late positivity:

The amplitude of the P600 has also been shown to be sensitive to the degree of syntactic "fit" of a phrase into a sentence. This may be analogous to the way in which the N400's amplitude has been shown to reflect the "semantic fit" of a word. Osterhout, Holcomb, and Swinney (1994) documented the ERPs to sentence-final phrases in four different types of sentential contexts. In two of the contexts, the preceding verb was either intransitive or "[could] be used either intransitively or transitively but shows a statistical bias towards intransitive usage"; both of these sentence types are easy to interpret syntactically. The examples they used are "The doctor hoped the patient was lying." (intransitive) and "The doctor believed the patient was lying." (biased towards intransitive). More difficult to parse were sentences with transitively-biased verbs that could also be used intransitively, as in their example, "The doctor charged the patient was lying." Sentences with verbs that are always transitive -- "The doctor forced the patient was lying." -- are purely ungrammatical, and cannot be parsed.

The ERPs to the two latter types of sentences clearly demonstrate the effects characteristic of the P600. Relative to the easy-to-parse sentences, the phrase in the entirely ungrammatical context generates a very large positivity. The phrase in a context which is grammatically permissible, but biased towards another kind of parsing, also generates a positivity -- but one which is much smaller in amplitude.

Returning to the notion of grammatical word classes, it should be noted that the classes are defined by the different sets of syntactic roles words can play. Nouns in English do not project themes, nor do they assign case, demonstrate tense, or otherwise behave syntactically like English verbs. If one interprets the P600 as "the ERP component sensitive to syntactic processing", then one might predict a difference in P600-epoch ERPs to nouns and verbs (and, by extension, to other syntactic categories of words: prepositions, for example). Just as the N400 is sensitive not only to semantic anomalies but to various semantic features of a word in its context, the P600 is sensitive to the kind of syntactic handling a given word evokes.

There is a question to be asked regarding any ERP seen in response to a linguistic task: is the effect purely and solely linguistic, or is it merely an example of a more general type of brain response to some more general class of stimuli?

There are a number of ERP effects which have been associated with non-linguistic stimuli. For example, the N2 components (peaking around 260-280 msec) has been shown to be sensitive to task relevancy, and orientation, spatial frequency, and location of visual features (O'Donnell et. al., 1997); the P300 varies in response to attended and/or task-relevant stimuli, for example in detection of unexpected notes in a tune (Paller et. al., 1992). (For a review, see Hillyard & Picton (1987)).

Deacon et. al. (1991) attempted to contrast tasks in which subjects discriminated among word stimuli on the basis of size, and another where they discriminated on the basis of semantic category. N2 and N400 were elicited only by their respective tasks, and were similarly distributed across the scalp, and similarly sensitive to changes in critical-stimulus probability. The N400 peaked about 80 milliseconds later.

Osterhout et. al (1996) used upper-case words (known to elicit P300-like waves) and number disagreement (known to elicit P600's) and combinations thereof to examine whether these two effects were similar. Most striking is the finding that the two effects were additive -- implying that they came from different neural sources. Manipulation of the frequency of occurance and task-relevance of the upper-case stimuli had robust effects on the ERP's; manipulations of those aspects of the number disagreement did not.

Coulson et. al. (1998) succeeded in eliciting significantly different P600-type effects by using two different types of syntactic violations (pronoun case violations and verb agreement violations); they argue that this lack of identicality between the responses to the two types of anomaly indicates a similarity to the nonlinguistic P3. However, other differences known to affect ERPs (such as closed vs. open-class words and word frequency) also existed betwen the two stimuli sets; given that, it is difficult to be certain of the conclusion that these differences are due to underlying similarity between the P600 and non-linguistic P3's. They did, however, succeed in eliciting a significant effect for task probability (by using 80% vs. 20% contrasts, as opposed to Osterhout et. al.'s 60% and 20%).

In consideration of noun/verb differences, the linguistic/nonlinguistic distinction is a rather coarse one to make. The difference between nouns and verbs is itself an essentially linguistic phenomenon; it has no real analogs (like frequency or task-saliency) outside of the linguistic realm.

A few studies have already attempted to address the difference between nouns and verbs in terms of evoked potentials. In an early series of electrophysiological studies, Brown et. al. (1973, 1976, 1979, 1980) found topological and timecourse differences in scalp EEGs recorded in response to nouns and verbs in sentential contexts. Their stimuli used homophonic nouns and verbs ("sit by the fire" (noun) versus "ready, aim, fire" (verb)) to avoid potential physical confounds in their stimuli. The differences were present only when the syntactic role of the word could be determined by prior context (so no differences were observed between "Fire is hot." and "Fire the gun."). Most interestingly, their results were robust across written versus spoken stimuli, and across two languages (English and Swiss-German).

Several issues, however, complicate the interpretation of their data. In their first experiment (1973), only two stimulus words -- "fire" and "duck" -- were used, in one noun context and one verb context apiece ("ready, aim, fire" / "sit by the fire"; "he'd better duck" / "a flying duck"). Stimuli were presented repeatedly, without regard for the effects of familiarity of the stimulus on the waveform. No baseline is given; it is difficult to determine what if any of the differences between the noun and verb forms are actually the continuing effects of the differences in preceding contexts. From a syntactic point of view, three of the stimuli are syntactically complete sentences, and one is a bare noun phrase; these are not linguistically comparable events. And no systematic analysis was done across classes to determine what about the elicited waveforms might be specifically related to the stimulus' grammatical class.

In the 1976 study, the problem of preceding context was solved by presenting only one stimulus ("lead", in the context of "it was lead", where subjects were told that this meant either "the horse was lead" or "the metal was lead"). A 1979 study added "an unintelligible sound sequence with suggested noun or verb meanings" -- i.e., a pseudoword, for which the subjects were told to imagine noun or verb meanings. The 1980 study added Swiss-German homophones. Similar differences were found in all the studies, although again the interpretations possible at the time were somewhat limited.

Despite the remaining issues with the methodology, the evidence that there is a difference of some kind makes the noun/verb distinction seem a very promising one to pursue. Samar and Berent (1986) examined nouns, verbs, and words with both noun and verb meanings, using a priming word that was either a posessive pronoun (which would, in an actual phrase, generally be followed by a noun) or a nominative pronoun (generally followed in a natural sentence by a verb). Subjects saw pairs such as "we" / "job" (unprimed noun), "we" / "eat" (primed verb), and "we" / "dance" (verb-primed ambiguous word). An overall comparison between nouns and verbs revealed generally increased negativity to nouns from around 200-350 milliseconds, and an increased negativity to ambiguous words from about 350-450. However, the sentential vs. nonsentential aspects of the stimuli were not analyzed ("we eat" and "we dance" are essentially valid, though incorrectly capitalized and punctuated, English sentences; "we job" is not).

In 1995, Preissl et. al. (apparently published again with some additional analysis as Pulvermueller et. al. (1996)) compared 50 concrete German nouns and 50 German verbs referring to motor actions, matched for length and frequency. (In German, unlike English, nouns and verbs tend to be quite category-bound, and homophones with both noun and verb meanings are rare.) Subjects rated motor and visual associations for all the words; this set of verbs was significantly more motor-related, and the nouns, more vision-related. EEG was collected as subjects performed a word/pseudoword discrimination task, and significant differences were noted during the P200 epoch, with verbs being more positive-going. They also did some topographical analysis of their findings, revealing positive-going responses to verbs over a large cortical surface, which they interpret as revealing of sources outside the classical language areas, which they interpret as reflective of the different kinds of associations (visual vs. motor) of the two kinds of stimuli they present.

However, their use of specifically "concrete German nouns" and "verbs referring to actions" can be taken as an inherently confounding -- are the differences observed due to the syntactic categories of the words in question, or because of their particular semantic connotations? There can, of course, also exist verbs that have strong visual associations and weak motor ones ("darken", for example), and nouns with strong motor and weak visual associations ("heaviness" and "flexibility", perhaps). An interesting variation would be to present sets of such words in contrast to the stimuli this experiment used.

An inspection of their figures appears to reveal an additional late positivity in response to verbs:

However, as the latest epoch they report was 260-350 msec, they do not report whether or not this difference was statistically significant, or speculate upon its possible theoretical implications.

Koenig and Lehmann (1996) used ERP "field maps" to document differences in brain responses to length- and frequency-matched nouns and verbs in German. Their single-word reading paradigm did elicit reliably nonidentical maps between subjects. But since little other work has been done to define the brain's response to syntax or semantics in terms of field maps, it is difficult to interpret the meanings of those differences in terms of their part in language processing.

Osterhout, Bersick, and McKinnon (1997) compared ERP responses to words in both a coherent, multisentence text context and a randomized version of that text. They found no significant differences between nouns and verbs in either case. However, their noun and verb stimuli were not intended to be matched for direct comparison. They used both single- and dual-category words in both classes. This could be confounding, especially in the randomized presentation; for example, "stretch", "search", and "stop" were coded as nouns; "photograph" and "rush" as verbs. Many of the verbs were definitely marked as such by their bound morphology (the suffix "-ed" appeard on over one-third; "-ing" on about ten per cent), but 54 (50% of unique verbs) were based on dual-category roots. English nouns generally involve less morphology and agreement phenomena than verbs do, and relatively few were marked -- but 35% also had verb frequencies listed in Francis & Kucera (1982). While sentence position was well-matched between nouns and verbs (mean position of 6.7 for nouns and 6.5 for verbs), position within phrase structures was variable -- English complicates this factor by allowing nouns to modify other nouns in a single phrase, so nominative cases such as "pilot" and genitives such as "flyer's" were both counted as nouns; the phrase "woman pilot" counted as two nouns. Furthermore, certainly in the continous-text case and to a lesser extent in the randomized one, repetition and priming factors probably reduced N400 amplitudes in some cases. The word "plane" accounted for 14 (7%) of the nouns presented, and only 63 (32%) of the nouns were unique; 97 (52%) of the verbs presented were unique, and the most frequent verb ("flying") appeared 7 times, accounting for about 4% of the (non-auxiliary) verbs. Also, several of the words ("radio", "mind", "claim") appeared under both codes, and a few of the "verbs" ("stormy", "skilled", "aware") appear to be more likely as adjectives. It may be that while in such naturalistic settings, brain response to nouns and verbs is indistinguishable, more subtle differences may be revealed when specific factors are more directly controlled.

In summary, the specific distinction between nouns and verbs has not previously been addressed in a way that attempts to control for other factors known to affect ERPs -- word frequency, sentential context, repetition, semantic associations, and/or physical characteristics of the stimuli.

The purpose of this research has been to use ERP techniques to examine brain responses to nouns, verbs, and pseudowords, in fully sentential contexts and a few cases of semantic and syntactic disagreements. Attempts were made to control such potentially obfuscating factors as preceding context, word frequency, and physical characteristics of the stimuli. Such data may demonstrate some differences and similarities between these syntactic categories, and enlighten the cognitive mechanisms underlying some linguistically-sensitive electrophysiological effects.

The primary hypothesis to be tested is, when all else is accounted for, do nouns and verbs elicit different kinds of ERPs? Do they do so regardless of other factors, or can other aspects of the words, such as frequency, overshadow any differences due to word class?

Subsidiary hypotheses include: If nouns and verbs are different, do those differences resemble any other ERP effects, for example the P600 class of effects associated with syntactic work? Is the P600 sensitive to syntactic phenomena such as word class, even in the absence of syntactic anomaly?

The experiments reported here investigate these questions from several different angles. First, an arbitrary assortment of nouns and verbs are compared, along with semantic anomalies and pseudowords in noun and verb sentence positions. Second, an experiment using frequency-matched, physically-identical nouns and verbs will compare the two syntactic categories while controlling for other factors known to affect ERPs. Third and fourth, the well-controlled sets of nouns and verbs are examined in contexts of semantic and syntactic anomaly.

These experiments represent an attempt to compare nouns and verbs directly, using an on-line, noninvasive physiological measurement of normally-functioning subjects, and using full natural language sentences.

Subjects were volunteers recruited from the community, between the ages of 18 and 35. Subjects in Experiment I received an ice cream cone for their participation; all subjects who so desired received a printout of their own ERPs. All had at least entered college; some had finished, and some had done postgraduate study as well. Subjects gave informed consent and were assessed for their handedness, native fluency in English, and freedom from known reading, learning, or neurological disorders. Subjects were tested in a single session that lasted approximately 2 hours, during which they were seated in a comfortable chair in an isolated room.

Each trial consisted of the following events: A fixation cross appeared for 500 msec, after which a sentence was presented, one word at a time. Each word appeared at the center of the screen for 350 msec, followed by a blank-screen interval of 300 msec, followed by the next word. The last word in each sentence was marked with a period, and followed by an 1150 msec blank-screen interval, after which a prompt appeared that said "YES or NO". Subjects would then be able to rest between trials, and would press a button when they were ready to see the next sentence. Which hand pressed the button was counterbalanced between trials. In some experiments a sentence acceptability task dictated which of two buttons to press; in others only one button was available, and subjects were instructed simply to read for comprehension.

English word frequencies were referenced from Francis and Kucera (1982). Highly wordlike pseudowords were constructed by normalizing overlapping trigram frequency to a standard English distribution. Stimulus sentences were generated by a native English speaker.

Continuous EEG was recorded from 13 scalp sites using tin electrodes attached to an elastic cap (Electrocap International). Electrode placement included the International 10-20 system locations (Jasper, 1958) over homologous positions over the left and right occipital (O1, O2) and frontal (F7, F8) regions and from the frontal, central, and parietal midline locations (Fz, Cz, Pz). In addition, several non-standard sites over posited language centers were use, including Wernicke's area and its right hemisphere homologue (WL, WR: 30% of the interaural distance lateral to a point 13% of the nasion-inion distance posterior to Cz), posterior temporal (L41, R41: 33% of the interaural distance lateral to Cz), and anterior temporal (L22, R22: midway between F7/F8 and T3/T4). Blinks and horizontal eye movements were monitored by means of an electrode below the left eye and an electrode to the right of the right eye. All the above were referenced to an electrode placed over the left mastoid process and amplified with a bandpass of .01 -100 Hz (3 dB cutoff) by a Grass Model 12 amplifier system. Activity over the right mastoid was actively recorded on a 16th channel.

Continuous analog-to-digital conversion of the EEG and stimulus trigger codes was performed by a Data Translation 2801-A board and a 486-based computer at a sampling frequency of 200 Hz. Epochs were comprised of the 100 msec preceding and the 1180 msec following presentation of individual words in the sentences. Trials characterized by excessive eye movements or amplifier blocking were removed from the analysis; subject runs with >30% trials rejected were removed entirely. The ERPs were quantified as the mean voltage within a latency range following presentation of words of interest relative to a baseline of activity comprised of the 100 msec prior and 50 msec subsequent to presentation of the words of interest.

Analyses of variance were performed on mean amplitude within four time windows: 50-150, 150-300, 300-500, and 500-800 msec post stimulus onset, with additional analyses as noted for some experiments. These epochs were chosen because they roughly correspond to the latency ranges of the N1, the P2, the N400, and the P600 effects respectively. (Since the total time between the onset of one stimulus and the onset of the following word was 650 milliseconds, the next word will actually have appeared on the screen during the last epoch; however, observation of the evoked waveforms reveals that P600-related effects still persist throughout that time.) Data from lateral and midline sites were analyzed separately in order to allow for quantitative analysis of hemispheric differences. The basic design for midline sites involved repeated measures on the factors of sentence type (condition) and electrode site; for lateral sites, sentence type (condition), electrode site, and hemisphere. Factors included syntactic word category (nouns vs. verbs) and word type (depending on the experiment, either at two levels: semantically appropriate versus inappropriate words, or three: appropriate, inappropriate, and pseudowords). Planned comparisons or simple effects analyses (Keppel, 1982) were made between nouns and verbs of each word type (i.e., good nouns versus good verbs), and between the different word types as well (for example, syntactically appropriate versus inappropriate nouns.) A general threshhold for "significance" was set at p=.05; given the number of comparisons made, a Bonferroni correction for familywise error was made, setting the significance for two-way contrasts at .04 and for three-ways at .035. Some additional peak-amplitude and time-frame adjusted analyses were performed post-hoc upon individual data sets to further explicate the phenomena observed.

This experiment contrasted the brain's responses to three types of words (real and appropriate words, semantically anomalous words, and pseudowords) in two syntactic positions within the sentence (noun and verb).

If nouns and verbs are really being handled differently by the brain, then this could be reflected in the ERPs generated while reading them. The amplitude or latency of certain components might be different between them, or the scalp distribution of electrical activity, or entirely different components might be visible.

Using anomalous and pseudowords in addition to real ones also allows one to examine the question of error handling in noun and verb contexts. The N400 component has long been well-known as a response to semantic anomaly, but a close reading of the appendices of that literature reveals that most of the stimuli used have been nouns in object position; perhaps verbs will be shown to behave differently.

Pseudowords present a particularly interesting kind of anomalous stimuli for two reasons. First, they are always one hundred per cent semantically unexpected; their Cloze probability is guaranteed to be absolute zero. Additionally, their syntactic properties are also one hundred per cent unknown; there is nothing in the lexicon to tell the reader whether the pseudoword is even syntactically appropriate to its noun or verb context. This provides an additional kind of contrast with the real words of both semantically expected and unexpected types.

Twenty subjects participated in this experiment. They were instructed to simply "read for comprehension", were warned that "not all of the sentences would make sense", and were to push a button when they were ready for the next stimulus to appear.

One hundred and fifty "sentence frames" were constructed. From each frame, five types of stimulus were generated:

(1) The skiers race down the mountain at sixty miles an hour.

(appropriate verb) (appropriate noun)

(2) The skiers bark down the mountain...

(semantically anomalous verb)

(3) The skiers spoo down the mountain...

(pseudoword verb)

(4) The skiers race down the fossil...

(semantically anomalous noun)

(5) The skiers race down the spoo...

(pseudoword noun)

Note that these word types have the following properties with respect to the sentences in which they appear:

appropriate (semantically and syntactically appropriate)

semantically anomalous (semantically inappropriate; syntactically appropriate)

pseudowords (i.e., letter-strings that are pronounceable and follow standard English graphemic rules, but do not actually happen to be real English words; neither semantic nor syntactic information about them can be found in the lexicon).

Critical words were matched over all frames for linear sentence position (this was achieved by exploiting the SVO order typical of English sentences, and placing critical nouns in both subject and object positions), and for physical characteristics such as length and word shape.

Five stimulus lists were generated. Each list contained one version of each frame, with thirty sentences of each stimulus type in each list, plus an additional sixty well-formed filler sentences, in random order. Each subject viewed one list, and each list was viewed by four subjects; two pressing the right button for YES, and two pressing the left.

Figures 7 and 8 plot ERPs to the three word types in noun and verb positions respectively:

Figure 7. appropriate, anomalous, and pseudoword nouns

In Figure 7, the appropriate nouns ("moon") can be seen to have evoked a fairly small N400 component, while the anomalous nouns ("potato") and pseudonouns ("spoo") evoked larger ones, quite similar to one another. After about 500 milliseconds, the two real-word cases overlapped to a large extent, while the pseudowords evoked a large and long-lasting positivity. In the verb conditions of Figure 8, there was much less difference among the conditions in the N400 epoch (though the appropriate verbs (such as "howl") were perhaps a bit less negative-going at some sites, and the pseudoverbs ("spoo") the most). In the P600 epoch, the anomalous verbs ("write") were distinctly more positive-going than the appropriate ones, and the pseudowords are even more so.

In the 50-150 millisecond epoch, a significant interaction between condition and electrode site was observed for nouns at midline sites (F(4,76)=3.15; p<.02) and by condition between hemispheres (F(1,11)= 6.59; p <.03) at lateral sites; for verbs, the interaction between condition and electrode site was again significant (F(4,76)=3.74; p<.01) at midline sites, as was the analysis by condition alone (F(1,11)=7.73; p<.02) at lateral sites. These very early differences appeared as a small increase in negativity for inappropriate nouns, and perhaps an increase in positivity for acceptable verbs. Given the early appearance of these effects, and the fact that preceding contexts were identical, the effects may be explained by physical differences in the stimuli presented in the different conditions. In the 150-300 millisecond epoch, overall ANOVAs revealed no significant differences.

The differences of the most theoretical interest involved mean amplitude between 300-500 milliseconds (the N400 epoch) and 500-800 milliseconds (the epoch of the P600). ANOVA for mean amplitudes for nouns in the 300-500 millisecond epoch revealed a significant difference by condition (F(2,38)=4.52; p<.02 at midline sites; F(2,38)=4.73; p<.02 lateral); of the pairwise comparisons, the only significant one was between appropriate and semantically anomalous nouns (F(1,19)=11.75 p<.01 midline; F(1,19)=12.81; p< .01 lateral). Between real and pseudoword nouns, significance was approached at lateral sites (F(1,19)=3.23; p<.09). This could be seen in the waveform as a classical large, central-parietal N400 effect. However, the effect was entirely insignificant for the various verb conditions at midline sites (F(2,38)=.6529; p > .5 for overall ANOVA). Verbs produced a visible N400-like effect only at left frontal sites, resembling the syntax-related LAN (overall ANOVA for lateral sites bordering on significance with F(2,38)=3.08; p<.06). Of the planned pairwise comparisons here, real versus anomalous verbs also bordered on significance after Bonferroni correction (F(1,19)=4.96; p<.04), as did real versus pseudoword verbs (F(1,19)=4.49; p<.05). Anomalous and pseudoword verbs did not differ significantly (F(1,19)=.82, p=.38).

In order to home in on the differences in the N400 component between nouns and verbs, an additional analysis was done, comparing peak-to-peak differences in the 350-450 millisecond range. For nouns, the differences were once again significant. Overall ANOVAs were F(2,38)=4.633; p<.02 at lateral sites and F(2,38)=4.36; p<.02 midline; pairwise between real and anomalous nouns F(1,19)= 9.12; p<.01 lateral and F(1,19)=10.34; p<.01 midline, between real and pseudoword nouns F(1,19)= 5.19; p<.04 lateral and F(1,19)=3.16, p<.10 midline. Anomalous nouns and pseudowords did not differ significantly in this analysis (F(1,19)=.0059, p=.94 laterally; F(1,19)=.68; p=.42 at midline), demonstrating that in terms of the N400, whether or not the word made sense in context was much more important than whether or not the word was actual English. For verbs, however, none of these peak-to-peak comparisons were significant.

In the 500-800 millisecond epoch, overall ANOVAs for nouns revealed significant differences by condition alone (F(2,38)=6.31; p<.01 at midline sites). Pairwise analysis revealed that this was not due to differences between real and semantically inappropriate nouns (F(1,19)=.0069; p=.93 midline), but to a pronounced positivity in response to the pseudowords (compared to real words: F(1,19)=6.07; p<.03 midline; compared to anomalous words, F(1,19)=11.78; p<.01 midline). For verbs, however, the picture looked very different: overall ANOVAs revealed highly significant differences (by condition alone: F(2,38)=9.17; p=.001 at lateral sites, F(2,38)=11.62, p=.0001 midline; by condition and electrode site at lateral sites: F(8,152)=2.47; p<.02; by condition, electrode site, and hemisphere: F(8,152)=2.18, p<.04). All pairwise comparisons were also significant (real vs. anomalous verbs: F(1,19)=4.60; p<.05 lateral, F(1,19)=5.37, p<.04 midline; real vs. pseudoword: F(1,19)=15.25, p=.001 lateral, F(1,19)=18.88; p<.04 midline; anomalous vs. pseudoword: F(1,19)=5.54; p<.03 lateral, F(1,19)=7.41, p<.02 midline). In the figure this appeared as a late positivity for anomalous words, and a similarly shaped but larger positivity for pseudowords.

Figures 9, 10, and 11 plot nouns versus verbs for the three word types:

Note that the anomalous and pseudoword plots overplot for both nouns and verbs. The differences in N400 and P600 effects seen in figures 8 and 9 appear to have been due to differences in the ERP responses to appropriate nouns and appropriate verbs (F(1,15)=5.32; p<.04 at midline sites during the 500 to 800 msec epoch) -- not the responses to anomaly. Although the differences between the good nouns and verbs were visually apparent during the N400 epoch as well, they were not statistically significantly so (at midline sites F(1,15)=1.29 and p=.27; lateral F(1,11)=4.21 and p=.52; all the p-values for interactions were >.45.) No other contrasts between these pairs were significant.

At first glance, the difference between appropriate nouns and verbs in Figure 7 seemed to be simply monophasic. However, when overplotted with the anomalous words, a more complex pattern emerged:

Appropriate nouns demonstrated a smaller N400 amplitude central-parietally, relative to semantic anomalies and verbs. Appropriate verbs had an entirely different effect: they demonstrated less positivity positivity from 500 to 800 milliseconds post-stimulus-onset, relative to the anomalies and the appropriate nouns.

This experiment yielded several interesting new observations. Pseudowords, which have no syntactic information associated with them in the lexicon, gave rise to large positive-going waves in the late (500-800 msec) epoch associated with P600 effects. Pseudowords elicited identical waveforms regardless of their syntactic position, and semantically inappropriate words elicited similar waveforms regardless of both their position and their category. But there were electrophysiological differences between appropriate nouns and verbs.

Pseudowords elicited N400 effects comparable to those elicited by unexpected real words, as well as large late positivities. The N400 effect fits in well with theories that relate N400 amplitude to semantic priming; if the word is a new one, its Cloze probability is zero and its relationship to previous semantic context nil, and a large N400 is to be expected. The late positivity resembles the Syntactic Positive Shift; its appearance may be explained by the fact that pseudowords, having no entry in the lexicon, can provide no syntactic information to aid in the ongoing parsing of the sentence.

The N400 effect elicited in response to semantically unexpected nouns replicated findings of many other experiments, though (when first presented as a poster in 1994) it was the first to do so explicitly for words of a single syntactic class. The failure to evoke distinctive N400 effects for verbs was provocative, and seemed to be indicative of basic differences between the noun and verb classes. However, as several other factors -- such as word frequency -- were known affect N400 amplitude, further experimentation was needed to make the significance (if any) of this finding clear.

Appropriate nouns and appropriate verbs in sentential contexts evoked significantly different ERP waveforms in this experiment. Nouns were more positive-going, which is unexpected, given the indications in other research that (a) positivity is associated with syntactic work, and (b) verbs contribute towards more syntactic work than nouns. However, the nouns and verbs in this experiment were not controlled for other factors known to affect ERPs, such as frequency and sentence position, so the specific contribution of syntactic category to this difference was not conclusively tested with this set of stimuli.

In anomalous and pseudoword cases, ERP response to nouns and verbs was essentially the same. This is evidence that supports the notion that the brain's way of handling unexpected or unknown words is similar in the cases of these two syntactic categories.

It is possible that the differences between nouns and verbs seen in Experiment I were due to some uncontrolled factor, such as physical differences or differences in frequency between the particular sets of nouns and verbs that were used. Word frequency can be controlled for with reference to studies done on that subject. And most conveniently, English provides a range of words that have both noun and verb usages, so all physical factors may be controlled by means of using the same words in their noun and verb forms. Both of these features were controlled in Experiment II, by contrasting physically identical, frequency-matched nouns and verbs.

(Although not directly examined, this choice of stimuli may also help to avoid the confound of the tendency of different syntactic categories to evoke different kinds of associations -- verbs may have generally more motor associations, and nouns more visual ones, but a word with both types of meaning -- such as "smile" or "dance" -- may be hoped to evoke similar sets of associations whether it appears in its noun or verb context.)

Data were collected for sixteen subjects. In the hopes of increasing subjects' attention to the stimuli, subjects were required to perform a sentence acceptability judgement for each sentence. This experiment was run concurrently with Experiment III.

(1) The children paint portraits.

(verb)

(2) The child's paint spilled.

(noun)

Noun and verb frequencies were normalized with reference to Francis and Kucera (1982) to have similar means and standard deviations in both usages. Mean noun frequency was 43.13, verb frequency 43.78; standard deviations were 68.27 and 85.80 respectively.

One hundred and twenty sentence pairs of the above type were generated. Four stimulus lists were generated, each containing thirty "critical verb" (as in (1) above) and thirty "critical noun" (as in (2)) sentences. Sentences were distributed across lists so that each list contained only one example of any critical word. Each list also contained sixty syntactically anomalous sentences (thirty with errors on the noun and thirty with errors on the verb; see Experiment III for details), and thirty well-formed "filler" sentences. Sentence types within the lists were randomly ordered. Each subject viewed one list.

Verbs generally appeared to be more positive-going than nouns throughout the length of the stimulus presentation.

Analyzed by epoch: No significant differences were found from 50-150 milliseconds. From 150-300 milliseconds, the difference bordered on significance (F(1,15)=3.54; p<.08 by condition at midline sites). No significant differences were found in the 300-500 msec epoch, though an interaction between condition and hemisphere approached significance (F(1,15)=2.95, p<.11), with the verb positivity appearing more prominent on the left. During the 500-800 millisecond epoch, the difference between conditions was highly significant (F(1,15)=8.38; p<.02 at midline sites; F(1,15)=4.77, p<.05 at lateral sites).

There were no significant interactions between condition and electrode site.

These data demonstrated a difference between nouns and verbs under highly controlled conditions. While it did not replicate the findings of Experiment I -- indeed, the difference seen was in entirely the opposite direction -- it actually fit better with findings in the literature that linked greater positivity to more syntactic work, and principles of linguistics that associated verbs with more syntactic work than nouns. Verbs in sentences assign case and determine the basic number of nouns needed in a sentence (as in transitive vs. intransitive verbs); in English at least, they also need to mark tense and number. In the example sentences, when one reached "paint" as a verb, one could note several syntactically-significant things: that it agreed in plurality with the preceding "children", that it was in the present tense, and that (being a transitive verb) some other noun (in this case "portraits") could appear in the sentence without any additional words or phrase structure to support it. When one encountered "paint" as a noun, though, little of this applied. The noun was marked as singular, but it demonstrated nothing about tense, and implied nothing about the syntactic structure of any phrase outside of its own (there are no "transitive" or "intransitive" nouns; nor are nouns capable of assigning Case).

It is also possible that two verb-related positivities are being evoked -- one in the P200 epoch, and a distinct one in the P600 epoch. This interpretation would fit well with Preissl et. al.'s (1995) suggestion that verbs give rise to a P200 by evoking motor-related associative activity in non-linguistic cortex. However, these stimuli were not controlled for motor or visual associations, which limits the interpretability of these data in that regard.

These findings were essentially replicated in the control sentences for Experiment IV, which placed the same words in different sentence contexts.

In general, nouns and verbs do different things syntactically. However, in English, they are both subject to at least one of the same syntactic requirements; they must both demonstrate number agreement with other elements of the sentence. As it happens, they even use the same marker -- a suffixed s (though it serves different purposes, marking the plural for nouns, and the third person singular for verbs). This allows us to investigate the differences between nouns and verbs, and syntactic agreement and disagreement, without sacrificing physical identicality in our stimuli.

This experiment added a syntactically anomalous condition to the noun/verb comparison of Experiment II. The contrasts of interest are twofold: Is the difference between a good noun and an unexpected one the same as the difference between a good and an unexpected verb? And do erroneous nouns and verbs produce the same or different responses from the brain?

This experiment was run concurrently with Experiment II, and subjects and task were the same.

(1) The children paint portraits.

(verb, agreeing)

(2) The child paint portraits.

(verb, disagreeing)

(3) The child's paint spilled.

(noun, agreeing)

(4) One of the child's paint spilled.

(noun, disagreeing)

(Sentences 1 and 3 are as in Experiment II.)

A pilot experiment indicated that on the word preceding the critical (underlined) ones, subjects had difficulty distinguishing between several similar written English noun forms: the s of nominative plurality, the 's of the singular genitive, and the s' of the genitive plural. (This would, of course, only be exacerbated by these forms' identicality in spoken English.) Of course, determining these features -- the case and number of the noun -- are critical to determining the appropriate class and markers on the critical word! So in this experiment, all stimuli used had distinctive irregular forms -- such as child, children, child's, children's in this example, or pronouns such as she, her, they, their-- so that the constraints on the critical word's form would be easier for subjects to perceive.

One hundred and twenty sentence pairs of the above type were generated. Four stimulus lists were generated, each containing thirty of each of the four types of sentences described above. Sentences were distributed across lists so that each list contained only one example of any critical word. An additional thirty well-formed "filler" sentences were presented as well. Sentence types within the lists were randomly ordered. Each subject viewed one list.

A separate analysis was done comparing appropriate nouns and verbs directly (see Experiment II).

This experiment could be considered as a two-factor design, crossing syntactic word category (nouns vs. verbs) with agreement (agreeing vs. disagreeing). Figure 14 plots agreeing versus disagreeing critical words:

Inspection of Figure 14 reveals that ERPs to disagreeing words were slightly more negative-going than ones to agreeing words from about 50-500 msec; this effect was largest over anterior and lateral sites (grammaticality by electrode site, 50-150 msec: F(2,30) = 4.68, p < .02; laterally, F(4,60) = 3.83, p < .02; 150-300 msec: lateral, F(4,60) =3.83, p<.05; 300-500 msec: lateral, F(4,60) = 3.25, p< .05). This effect appeared to fall into the LAN category, visible as an increase in a negative-going peak about 400 msec post stimulus onset, largest at left anterior sites. Beginning at about 500 msec and for the rest of the epoch, ERPs to disagreeing words became increasingly positive-going relative to agreeing ones (500-800 msec: midline, F(1,15)=6.01, p < .1; lateral, F(4,60)=3.25, p < .05); this has the look of a late syntactic positive shift.

Figure 15 plots ERP responses to nouns and verbs, lumping together syntactically agreeing and disagreeing occurances. ERPs to verbs were more positive-going throughout the entire recording epoch. This included the P2 epoch, particularly over anterior regions (word class by electrode site, 150-300 msec: midline, F(2,30) = 3.35, p < .05; lateral, F(4,60) = 3.40, p <.02) and in the left hemisphere (word class by hemisphere: F(1,15) = 4.52, p = .05). There was also a late positivity associated with verbs, beginning at about 400 msec and persisting through the rest of the epoch, and significant at midline sites (500-800 msec: midline, F(1,15) = 11.33, p < .01).

A question of primary interest is whether number violations on nouns and verbs elicit similar or dissimilar responses. A dissimilar response would be evidenced by significant interactions between word class and grammaticality, possibly involving additional interactions such as hemisphere or electrode site. Only three of the interactions in this set approached statistical reliability: word-class by grammaticality by electrode site at lateral sites from 150-300 msec (F(4,60) = 3.49, p < .05); word-class by grammaticality at midline sites from 300-500 msec (F(1,15) = 3.24, p<.1); word-class by grammaticality by electrode site from 500-800 msec (F(2,30) = 2.95, p<.07).

That syntactic anomaly produces late positive-going activity has been widely documented. Examining the data in terms of simple effects analyses, in this experiment, number disagreement on nouns was marked by a long-lasting and widely-distributed positive shift, as shown by Figure 16:

Differences between nouns agreeing and disagreeing for number were statistically significant even from the earliest epoch (at lateral sites, interaction between condition and electrode site (F(4,60)=3.24, p<.02; interaction between condition, electrode site, and hemisphere (F(4,60)=2.49, p=.05; at midline sites, interaction between condition and electrode site (F(2,30)=3.54; p<.05); this was, however, reversed in polarity on various parts of the head, and more likely reflected baseline artifacts than actual brain responses to the syntactic anomaly of the current stimulus word. From 150-300 msec, the only significant effect was an interaction between condition and electrode site at lateral sites (F(4,60)=4.95; p<.01), reflecting a larger P2 peak to disagreeing nouns at right rear sites, and contrariwise to appropriate nouns at the F8 electrode. From 300-500 msec, there was an interaction between condition and electrode site siginificant at lateral sites (F(4,60)=3.55, p=.01) reflecting a smaller N400 peak to disagreeing nouns at right rear sites. The largest and most broadly distributed effect was the late positivity seen during the 500-800 msec epoch (by condition alone: F(1,15)=5.66, p=.03 at lateral sites; F(1,15)=6.09, p<.03 at midline sites; interaction between condition and electrode at lateral sites, F(4,60)=6.41, p<.001). (This was also the only epoch in which a significant effect appeared in response to condition alone.)

The same general trend of late positivity in response to anomaly was apparent for verbs, but was a little differently distributed, being more pronounced at midline sites and having a somewhat later onset, with a very late peak, as can be seen in Figure 17:

During the P600 epoch, condition and electrode interactions were significant; (F(4,60)=4.92; p<.01 at lateral sites; F(2,30)=5.58, p<.01 at midline sites). There appeared to be a left-hemisphere increase in N400 peak for anomalous verbs, but this effect was not statistically significant (F(1,15)=1.194, p=.18).

Figure 18 plots disagreeing nouns versus disagreeing verbs:

The primary finding of the difference between nouns and verbs was seen again in the comparison between violating nouns and violating verbs -- verbs were significantly more positive-going during the 500-800 msec epoch at midline sites (F(1,15) = 6.7454, p=.02). There were also significant interaction effects between electrode site and syntactic category, extending throughout the entire ERP (F (2,30) = 3.21, p=.05 at midline sites from 50-150 milliseconds; F(4,60) = 3.92, p<.01 lateral and F(2,30) = 4.00, p<.03 midline from 150-300 msec; F(4,60) = 2.83, p<.04 lateral from 300-500 msec, and F(2,30) = 2.64, p<.09 midline from 500-800 msec), and reflecting effects largest at posterior sites.

Since the syntactic failure of number agreement was essentially the same in both violating cases, I predicted that the difference waves for (1 - 2) and (3 - 4) might be the same; these were not found to differ significantly (p-values for all analyses > .3).

These results demonstrated that the same syntactic manipulation -- number agreement versus disagreement -- evoked the same ERP response -- the late positive shift -- regardless of whether the critical word was a noun or a verb. Furthermore, the waves evoked by syntactically anomalous nouns and verbs preserved the same difference between well-formed nouns and verbs seen in Experiment II -- verbs were more positive-going, especially in later epochs. Indeed, ill-formed nouns and well-formed verbs evoked remarkably similar-looking waves, as can be seen in Figure 19:

This experiment compared physically identical nouns and verbs in semantically expected and unexpected contexts. Again, the contrasts of interest were: would good nouns and good verbs evoke similar or different ERPs? Was the difference between a good noun and an unexpected one the same as the difference between a good and an unexpected verb? And would erroneous nouns and verbs produce the same or different responses from the brain?

Data were collected for sixteen subjects. Subjects were required to perform sentence acceptability judgements on each sentence read.

As Experiment III manipulated the syntactic validity of the critical words in the sentence, Experiment IV placed the critical words in semantically acceptable or violating contexts:

(1) The wolf howls at the moon.

(verb, expected)

(2) The work howls at the moon.

(verb, anomalous)

(3) The wolf's howls awakened us.

(noun, expected)

(4) The work's howls awakened us.

(noun, anomalous)

Again, one hundred and twenty sentence frames were generated, each yielding one sentence of each type per frame. Four lists were created, each containing thirty sentences of each of the four types, and an additional thirty well-formed "filler" sentences. Order of types was randomized in the lists. Each subject saw one list.

Figure 20 plots semantically expected words versus unexpected ones, collapsing nouns and verbs together. A small (~1 microvolt) increase in negativity, starting around 200 milliseconds and persisting through the N400 (300-500 millisecond) epoch is apparent in responses to the anomalies. This is smaller than most N400 effects of semantic anomalies reported in the literature, which tend to be in the range of 2-5 microvolts. The difference between syntactically expected and anomalous cases was reliable over lateral sites between 300 and 500 msec (F(1,15) = 6.65, p < .03), but not midline sites (F(1,15) = 1.65, p = .21).

Figure 21 plots responses to nouns and verbs, collapsed across anomalous and non-anomalous cases. Verbs were generally slightly more positive-going from about 200 milliseconds on. Although these effects were similar in character to those seen in Experiments II and III, they were smaller in amplitude, and not statistically robust (main effect for word class: midline sites, 500-800 msec, F(1,15) = 3.07, p=.1; lateral, F(1,15) = 0.23, p > .6; word class by electrode site: midline, F(2,30) = 2.48, p = .1; lateral, F(4,60) = 4.37; p < .01).

Turning to the planned pairwise comparisons, good nouns and good verbs were found to evoke ERP's of similar general shape, but with greater positivity associated with the verbs:

These stimulus types differed significantly throughout the trials -- F(4,60)=3.2827, p<.02 during the 50-150 msec epoch (as an interaction between condition, hemisphere, and electrode position, and visible only at left frontal and temporal sites); F(1,15)= 4.5077, p=.05 at midline sites during the 150-300 msec epoch; F(1,15)=6.9203, p<.02 at midline sites from 300-500 msec; and F(1,15) =4.50, p=.05 at midline sites from 500-700 msec. The very early effects were perhaps due to syntactic expectations set up by the differing classes of the preceding word in the sentence -- determiners lead to an expectation of a noun; nouns lead to an expectation of a verb. All these findings followed the familiar pattern of noun/verb differences, with the verbs more positive-going, and the differences generally most visible at central sites.

Expected and unexpected verbs were significantly different -- looking quite like the classical N400 effect, with anomalous cases showing a clear increase in the negativity of that component:

Significant differences were detected in the 150-300 millisecond epoch (by an interaction between condition and electrode site at lateral sites, F(4,60) = 3.32, p<.02), at midline sites in the 300-500 msec range (F(1,15) = 4.32, p=.05), and at lateral sites in the 500-800 millisecond epoch (F(4,60) = 2.73, p<.04 for the interaction between condition and electrode sites). (These epochs were selected a priori as "standard" for all experiments, and may not best characterize the actual components of interest under these particular conditions.)

Semantically unexpected nouns and verbs were not found to differ significantly from each other at all:

Interestingly and unexpectedly, good and bad nouns did not differ significantly from each other either:

(Midline, F(1,15) = 0.03, p=.86; lateral, F(1,15) = 0.87, p = .36.) This finding very likely accounts for the smallness of the effects seen when cases were collapsed in Figures 21 and 22.

This experiment confirmed that the N400 effect occured in response to semantic anomaly on verbs. It did not, however, demonstrate similar effects in response to semantic anomalies for nouns. This seems curious, given that an inspection of the appendices of the N400 literature reveals a predominance of nouns in the critical positions for those experiments. It may be worth noting that in both the anomalous and non-anomalous conditions of this experiment, the N400 components' magnitudes were quite large (~7 microvolts at Pz). Since in this experiment, the critical word generally appeared around the third word in the sentence, this fits in well with Van Petten & Kutas (1990)'s finding that words appearing early in a sentence elicit larger N400's than words appearing later. But when the critical words appeared in verb form, even though their distance into the sentence was positionally equivalent to the nouns (that is, also the being third word in the sentence), they were nonetheless deeper in in the syntactic sense: part of the second phrase (the verb phrase) rather than the first (as is the case with the subject noun). This syntactic depth may reflect more buildup of semantic constraint than a simple count of preceding words regardless of their syntactic influence on the sentence, and result in a significant decrease of the N400 of a critical word in the second phrase (i.e., the verb).

(It should be noted that verbs' general greater positivity relative to nouns in these experiments cannot be attributed to syntactic depth alone, as it was also observed in Experiment I, which used nouns in both subject (first phrase) and object (third phrase) positions; also, in all the other experiments, it was generally not apparent until after the N400's usual epoch.)

The experiments reported here comprised an attempt to discover what, if any, systematic differences exist in the brain's electrophysiological responses to nouns and verbs. To this end, they directly contrasted the brain's response to nouns and verbs in appropriate sentential contexts, and in conditions of semantic and syntactic violation. One experiment also documented the brain's response to pseudowords in the contrasting sentential contexts of nouns and verbs.

The primary findings of these studies are: that verbs evoked more positive-going ERPs than do nouns, that semantic and syntactic anomalies had similar effects on both nouns and verbs, and that pseudowords evoked both semantically- and syntactically-related ERP effects.

No electrophysiological differences were found between pseudowords encountered in noun positions and those encountered as verbs. Still, the brain's response to pseudowords relative to real words is, of itself, very interesting. Pseudowords have a long tradition of use in cognitive psychology, dating back to Ebbinghaus's 1885 experiments on memory (though Ebbinghaus did specifically try to avoid "meaningful" analyses of his pseudowords; that is, he tried to avoid forming semantic or syntactic associations with them). The presentation of pseudowords in sentential contexts is also of interest for its similarity to natural word learning; all words may be considered "nonsense" the first time we encounter them. But as Carey and Bartlett (1978) have demonstrated, a single "nonsense" encounter may be enough to allow us to recognize a new word when we encounter it a second time.

What does the brain do when it encounters an unknown word in the context of a sentence? According to these data, it generated both a large N400 component and a large late positive shift. The N400 was not observed to be any larger than that generated in response to a real but semantically unexpected word. The amplitude of the N400 appears to reach a maximum when Cloze probability becomes zero, and not to increase beyond that. That pseudowords should generate P600s is also not surprising, since no syntactic information can be retrieved about them from the lexicon, and the parser must attempt to go ahead without any help from that word.

Two aspects of the P600s evoked by pseudowords in these experiments may still be considered surprising. For one thing, they are generated even though it is, in English, quite easy to make good guesses about the syntactic nature of the pseudowords, given their contexts. Even in sentences loaded with pseudowords, as in the famous, "'Twas brillig, and the slithy toves did gyre and gimble in the wabe.", it is not a problem for the reader to realize, at the conscious level, that "gimble" is a verb and "wabe" is a noun. However, at the level of processing reflected by the electrophysiology of the P600, the failure is nonetheless apparent, and the P600's are large. Another surprising element may be that the P600's overlap for pseudowords in noun position (where little besides category-membership and some agreement might be expected of them), and in verb position (where tense and Case-assignment, as well as agreement phenomena and category, will fail). Perhaps the P600's in both cases are as large as P600's can be; the failure to retrieve them from the lexicon evokes a ceiling-level effect.

A pilot experiment not reported in detail here (the "vicka" or "Jabberwock" experiment) appeared to demonstrate that the P600s generated by real words with syntactically inappropriate bound morphology, pseudowords with appropriate morphology, and pseudowords with inappropriate morphology all generate similar P600s relative to real words marked appropriately:

A better-controlled experiment would be needed to determine the relative contributions of morphology and word-stem lexical information to the ERP components. But again, the general similarities in size of the P600s in these various conditions may point to a ceiling effect.

Experiment I appeared to demonstrate that ERPs to nouns were basically more positive-going than to verbs. However, the stimuli presented in Experiment I were not well-controlled; the nouns and verbs in question were chosen without regard for certain features known to affect ERP response, such as frequency. Stimuli for Experiments II, III, and IV were chosen to be physically identical and of matched frequency as nouns and verbs. Under these conditions, their findings indicate that verbs are generally more positive-going than nouns, particularly during the 500-800 millisecond epoch.

Specifically, Experiment II revealed that well-formed verbs were more positive-going in the P600 epoch than identical well-formed nouns. Experiment III showed that this difference was maintained when the nouns and verbs in question were syntactically violating as well. Experiment IV demonstrated that, when placed in the same numeric position early in a sentence, verbs (which are syntactically part of the sentence's second phrase) produce smaller N400 waves than nouns (which are part of the first phrase in the sentence).

The association of positivity with verbs fits in well with the model of late positivity as a marker of syntactic work. In all languages, verbs perform a number of syntactic duties not associated with nouns. For example, they assign case and project thematic roles. Consider the following sentences:

1) Susan smiles.

2) Susan wants a puppy.

3) Susan gave the puppy water.

*4) Susan smiles the puppy.

*5) Wants.

*6) Susan gave the puppy water food.

Each verb is associated with having a certain number of nouns (or, more properly, noun phrases) going along with it in sentences; having either too many or two few creates a syntactic error. If one wants to create a sentence with more noun phrases than the verb in question can assign Case to, one needs to put in an additional Case-assigning word such as a preposition: "Susan wants a puppy around the house." Nouns do not dictate the structure of the surrounding sentence in such a way in any language; no noun requires, for example, two verb phrases to form an acceptable sentence.

In English, some additional syntactic work is also visited upon verbs: they are marked for tense, for number, and for person. English nouns show very little syntactic marking; they demonstrate number, and a few pronouns also demonstrate person. It is possible that it is this language-specific feature underlies the additional positivity associated with the verbs in these experiments done in English. It would be interesting to see how the P600 behaves with regard to syntactic word category in languages that are different; for example the case-marked nouns of Japanese, or Chinese verbs which show neither tense nor agreement.

Regardless, the findings here demonstrate that the association of positive-going ERPs with syntactic work does not begin and end with anomalous or unlikely words. Additional positivity was observed to verbs that fit well into their contexts both semantically and syntactically, relative to physically identical and frequency-matched nouns. This may be seen as early as the P2 component (Experiment IV), but is most pronounced and reliable later on, particularly in the 500-800 millisecond epoch (Experiment II).

In general, anomalous nouns and verbs (or pseudowords in noun and verb position) seemed to elicit highly similar ERP responses (Experiments I, III, and IV). The ERPs to anomalous nouns and anomalous verbs were generally not found to differ significantly, either in semantically (Experiments I and IV) or syntactically (Experiment III) violating contexts. Nor were pseudowords presented in noun or verb position found to be different from each other (Experiment I). Number disagreement elicited P600 effects for both nouns and verbs (Experiment III).

Semantic errors on verbs were shown to elicit the N400 effect (Experiment IV). Interestingly, Experiment IV's semantically violating nouns did not elicit an enhanced N400 component, although such findings are well-documented in the literature, and generally quite robust. This was the case even though the nouns and verbs in question were located the same number of words into the sentence and were preceded by almost identical contexts:

The wolf howls ....

The work howls .... (verb; increased N400 amplitude)

The wolf's howls ....

The work's howls .... (noun, did not increase N400)

These data do not actually cast much doubt on the extremely robust and common finding that semantic anomaly evokes the N400 in the noun case. Most likely, the failure to find an N400 effect of significance in the good-vs.-bad nouns analysis owed mainly to the fact that the critical words appeared quite early in the sentences, before very much constraining context had appeared. Given that, the appearance of a clear and significant N400 response to anomalies on the verbs must be considered particularly noteworthy, and demands some explanation.

It seems unlikely that "The work's howls" has a similar Cloze probability to "The wolf's howls". Nor can the difference between nouns and verbs be due to the simple lengthof the preceding context -- both kinds of critical words appeared in the same numerical position relative to the beginning of the sentence. It may, however, relate to the syntacticdistance into the parse tree. The nouns in Experiment IV appeared in the first phrase of the sentence, and the verb in the second; in most of the literature, the critical words appear in the third (as object nouns). In Experiment I, which used nouns placed both in subject and object phrases, clear N400 responses to semantically violating nouns were observed.

Another hypothesis might be that nouns may semantically constrain the verbs that follow them more than genitive determiners constrain nouns. That is, it may seem less implausible for a "work" to have"howls" -- perhaps as a work of performing art? -- than to contemplate a "work" that actually "howls" itself.

These experiments' findings were generally well in line with literature that predicts that there should be differences between nouns and verbs. Experiments II, III, and IV, which compared identical, frequency-matched nouns and verbs, demonstrated that under those conditions, verbs evoke more late positivity than nouns do. Experiment III also demonstrated that responses to the syntactic anomaly of number agreement are similar whether evoked by a noun or a verb. And while Experiment IV did not succeed in evoking any significant response to syntactic anomaly in nouns, in its verb conditions, it demonstrated an N400 effect similar to that extensively documented (generally in primarily-noun-based stimuli) elsewhere in the literature.

There is an inconsistency between the noun-and-verb differences seen in Experiments II, III, and IV, and the opposite-direction difference seen in Experiment I, and also in the lack of any systematic differences seen in Osterhout et. al. (1997). This may be due to the particular nouns and verbs selected for the various experiments. English has more nouns than verbs, and nouns tend to be more frequent. Experiment I's effects could be interpreted as a persistent reduction of the N400 due to higher frequency of the noun population, and Osterhout et. al. (1997)'s lack of significant differences due to a wider range of frequencies, repetition and Cloze-probability effects, and the mixture of dual- and single-category words in both conditions.

Indeed, an alternative explanation of the findings of Experiments II, III, and IV is that the verbs' late positivity is specifically an effect of using dual-category words. If, for example, all English dual-category words were categorized as "nouns" at some first-parse lexical level, then the late positivity seen when they appear in their function as verbs may simply be the result of a syntactic mismatch and the need for reparsing.

An experiment specifically contrasting dual-category words (the "physically identical" nouns and verbs used in experiments II, III, and IV) with carefully association-, shape- and frequency-matched single-category nouns and verbs would be useful to resolve these open issues.

These data also shed some light on the cognitive processes that must underlie the electrophysiological behaviors of the ERPs.

In those experiments that contrasted physically identical, frequency-matched nouns and verbs, verbs generated more positive-going brainwaves, particularly in the peak of the N200 component and in the P600 epoch. This finding, when viewed in the light of linguistic theories that attribute more syntactic duties to verbs than to nouns, demonstrates the association of ERP positivity with syntactic work under normal, linguistically-acceptable conditions. The association of the P600 with syntactic anomaly is well-documented (and demonstrated again here in Experiment III), as is its association with syntactically difficult conditions such as "garden path sentences" (Osterhout & Holcomb, 1992, 1993). But even though there is nothing syntactically anomalous or difficult about the verb in the sentence "The wolf howls at the moon.", it still evokes greater positivity than the subject noun in "The wolf's howls awakened us." One may speculate that this positivity is in fact evoked by the same processes that underlie the SPS.

This leads to another interesting speculation -- that some of the differences seen here between nouns and verbs may be due to the phrase structure of the sentences at the point of the experiments' critical words. The difference in phrase structure is actually apparent on the word *preceding* the critical one. The preceding word was a noun in genitive form ("wolf's", or "his"), which would be followed by a nominative-case noun (the last word in the phrase) in the "critical noun" sentences; in the contrastic "critical verb" setences, the preceding word was a nominative noun ("wolf" or "he", and therefore also the end of a phrase).

Post-hoc analyses were done. The data are suggestive, though not entirely statistically reliable; in Experiments II, III, and IV, the nominative nouns (phrase enders) were more positive-going than the genitive forms (middle-of-the phrase words). In Experiments II and III (which were analyzed together) significance was not acheived at either midline or lateral sites during either the 300-500 or 500-800 msec epochs, though p < .2 for all four analyses. In experiment 4, midline sites were significantly different by sentence type during the 300-500 msec epoch (F (1,15) = 6.79, p<.02; during the 500-800 msec epoch, lateral sites were significantly different in an interaction between sentence type and hemisphere (F (1,15) = 4.34, p=.05). Given that the interstimulus interval in both experiments was only 650 milliseconds, these preceding-word effects may account for some of the very early (50-150 millisecond) effects seen in experiments I, III, and IV. Further investigation specifically regarding syntactically different occurances of nouns (such as the genitive/acting-like-an-adjective and nominative/ending-noun-phrase types seen here) would be useful in understanding this issue.

Another issue enlightened by these data is the difference between the P300 class of ERPs and the P600 or SPS. Verbs evoked significant positivity relative to physically identical nouns matched for frequency and sentence position. The differences between the two classes -- nounness versus verbness, Case-assignation, the casting of roles for other words in the sentence, agreement phenomena -- are purely syntactic. Neither nouns or verbs can be said to be more "task-relevant" to the process of reading a well-constructed, coherent sentence; both classses are vital.

Like most projects, this dissertation does not exhaust all possible investigation into the nature of the questions it addresses. Several issues remain obscured, and only further investigation can hope to address them.

These experiments were done only in English. Other languages are not necessarily like English, and some of the differences among languages may be critical to the kinds of phenomena measured here. For example, English verbs take a few kinds of bound morphology, indicating tense and number. English nouns take only number. One possible interpretation is that the additional positivity seen to the verbs in this experiment is that it is elicited by the parser's need to extract more kinds of syntactic information from the verbs than from the nouns. This confound could be addressed by adding experiments involving different languages, with different types and levels of syntax associated with different categories. For example, Chinese verbs take no bound morphology; Hebrew nouns show gender as well as number; Japanese nouns show case -- the variations across languages are many, and the investigative possibilities are vast.

One particular feature of English exploited by several of these studies is the existence of dual-class words. Not all languages have them, and few indeed allow the relatively free crossing of the open classes that English does. Single-class words may behave differently, or dual-class words may even behave differently in different languages. For example, it is possible to interpret the additional positivity seen in response to dual-class verbs to be an artifact of reclassifying a word that (despite having been balanced for noun and verb frequency in the Brown corpus) is basically a noun in the lexicon.