A Rose by Any Other Name: Long-Term Memory Structure and ...ling.umd.edu/~ellenlau/courses/ling848/Federmeier_1999.pdf · A Rose by Any Other Name: Long-Term Memory Structure and

Journal of Memory and Language41, 469–495 (1999)Article ID jmla.1999.2660, available online at http://www.idealibrary.com on

A Rose by Any Other Name: Long-Term Memory Structureand Sentence Processing

Kara D. Federmeier

Department of Cognitive Science, University of California, San Diego

and

Marta Kutas

Departments of Cognitive Science and Neuroscience, University of California, San Diego

The effects of sentential context and semantic memory structure during on-line sentence processingwere examined by recording event-related brain potentials as individuals read pairs of sentences forcomprehension. The first sentence established an expectation for a particular exemplar of a semanticcategory, while the second ended with (1) that expected exemplar, (2) an unexpected exemplar fromthe same (expected) category, or (3) an unexpected item from a different (unexpected) category.Expected endings elicited a positivity between 250 and 550 ms while all unexpected endings elicitedan N400, which was significantly smaller to items from the expected category. This N400 reductionvaried with the strength of the contextually induced expectation: unexpected, categorically relatedendings elicited smaller N400s in more constraining contexts, despite their poorer fit to context (lowerplausibility). This pattern of effects is best explained as reflecting the impact of context-independentlong-term memory structure on sentence processing. The results thus suggest that physical andfunctional similarities that hold between objects in the world—i.e., category structure—influenceneural organization and, in turn, routine language comprehension processes.© 1999 Academic Press

Key Words:sentence processing; categorization; event-related potentials; N400.

lvel-fo

heypata-n-ord

ntthe

con-heirthein-pleusede,

s tois

rge,g-

t isoutand

ten

doc13,anash,lpfu

paran

15.

At its heart, language comprehension invothe recruitment and integration of world knowedge stored in long-term memory. Consider,example, the following pair of sentences:

“Getting himself and his car to work on the neigh-boring island was time consuming. Every morning hedrove for a few minutes and then boarded the . . . ”

When asked, most individuals report that texpect the missing final word of this sentenceto be “ferry.” How do they come to that expection? None of the individual words in this setence pair is strongly associated with the w

This work was supported by a Howard Hughes Pretoral Fellowship to K.F. and grants HD22614, AG083and MH52893 to M.K. We thank Elizabeth Bates, SeCoulson, Phillip Holcomb, Jonathan King, David KirMarty Sereno, and two anonymous reviewers for hecomments and discussion.

Address reprint requests to Kara D. Federmeier, Dement of Cognitive Science, University of California, SDiego, 9500 Gilman Drive, La Jolla, CA 92093-05
E-mail: [email protected]; Fax: 619-534-1128.
469

s

r

ir

ferry. There are, in fact, a number of differevehicle names that could plausibly completesentence. Yet, the expectation is remarkablysistent across individuals. In order to create texpectations, readers must use information insentence context to build a cognitive modelvolving vehicles that can transport both peoand cars across water and that are likely to behabitually. It is their store of world knowledgcombined with this model, that allows readerthen determine that the vehicle in questionlikely to be a ferry and not an ocean liner, baairplane, or helicopter. Given how crucial lonterm memory is for language processing, isomewhat surprising that so little is known abhow information from memory is accessedused during on-line language processing.

Context Effects in Language

Off-line tasks make it clear that there is of

-

l

t-

sufficient information in a sentence context to

0749-596X/99 $30.00Copyright © 1999 by Academic Press

All rights of reproduction in any form reserved.

onenheninguchord

ara

ccuorth&

on-ro-rdss1;

oon0;ili-lex

ento

t besol

ff,

esetiontheen

vesis

ram-re

ral7).u

dutheesor-seo aes,

pe-u-0, a400rd

k inordthatt toth a

sthe

tionsnce

hisu-oze-vertas,

onon-in

i.e.,heee,

tords.mebuttes,w-9;nt,ns,msose

lermsralob-are

tions,

erst letet ord

470 FEDERMEIER AND KUTAS

significantly constrain guesses about what ccepts or even words are likely to be nextcountered. However, whether—and, if so, wand how—this information affects processon-line remains a hotly debated issue. Mpsycholinguistic research suggests that wthat are predictable in a sentence contextperceived and processed more rapidly andcurately than the same words when they oout of context or in incongruent contexts. Fexample, contextual information decreasesduration of readers’ eye fixations (EhrlichRayner, 1981; Morris, 1994; Zola, 1984). Cgruent contexts also facilitate the time to pnounce sentence-final or phrase-final wo(Duffy, Henderson, & Morris, 1989; Hess, Fo& Carroll, 1995; McClelland & O’Regan, 198Stanovich & West, 1983) and the speedword/nonword judgments (lexical decision)them (Fischler & Bloom, 1985; Kleiman, 198Schuberth, Spoehr, & Lane, 1981). This factation occurs even when the relatedness ofical items within congruent and incongrusentences is matched, suggesting that theserved increase in processing fluency cannoattributed solely to lexical priming, but involvinformation provided by the sentence as a wh(e.g., Duffy et al., 1989; Morris, 1994; Ratcli1987).

Electrophysiological results support thfindings and suggest that contextual informais used early and builds continuously overcourse of processing a sentence. The evrelated brain potential (ERP) technique involrecording at the scalp neural activity thattime-locked to a particular event. The neuactivity recorded is known to reflect the sumation of graded postsynaptic potentials, pdominantly from pyramidal cells of the cerebcortex (for review, see Kutas & Dale, 199The ERP technique provides a continuous, mtidimensional measure that can be recordeding natural language processing (withoutimposition of an additional task). It providmillisecond-level temporal resolution and infmation about the number and, in some calocation of the neural sources contributing tgiven task or condition (e.g., Rugg & Col
1995). (
--

se

c-r

e

s,

f

-

b-e

e

t-

l

-

l-r-

s,

An ERP component that has proven escially useful for the study of contextual inflences in language processing is the N40negative-going potential peaking aroundms after stimulus onset. Kutas and Hillya(1980b) first observed the N400 during a taswhich individuals read sentences word by wfor comprehension. Sentence final wordswere semantically anomalous with respecthe sentence context were associated wisignificantly larger negativity 250 to 600 mpost-stimulus-onset than were words that fitsentence context. Subsequent investigahave revealed that each word in a senteelicits an N400 and that the amplitude of tcomponent is highly correlated with individals’ off-line expectations as measured by “clprobability1” (Kutas & Hillyard, 1984) and decreases as contextual information builds othe course of a sentence (Van Petten & Ku1990).

The influence of contextual informationword processing has been most clearly demstrated for words that are highly predictabletheir sentence contexts (“best completions”;words with the highest cloze probability in tcontext). However, to a more limited degrcontextual information has also been foundaffect the processing of less predictable woWith behavioral techniques, for example, soresearchers find facilitation for unexpectedcontextually congruous words (e.g., Schwan1985; Stanovich & West, 1983); others, hoever, do not (e.g., Fischler & Bloom, 197Kleiman, 1980; Schwanenflugel & LaCou1988). In electrophysiological investigatiothese congruent but low cloze probability iteelicit N400 responses that are larger than thto higher cloze probability items but smalthan those to contextually incongruent ite(e.g., Kutas & Hillyard, 1984). Both behavioand electrophysiological studies have alsoserved facilitation for unexpected items thatsemantically related to the best comple(Kleiman, 1980; Kutas & Hillyard, 1984; Kuta

1 The cloze probability of a word in a given context refo the proportion of people who would choose to comphat particular sentence fragment with that particular w
Taylor, 1953).

elobeptKu

extto

ort

le-xts

erwetol &

entheroblewiti.eh aaciffycer

heron-lly

stsdeepngtonw

nceotobseforne

usein-y-thiselyw

se-ter-extortileof

utes-&ices-

e topo-ionsle,ith,resmedm-u-s etentof

rec-lso

ct tonge-

ngsrch

thehisses.y,ialings forings

in-d ces wes fert ectee

471LONG-TERM MEMORY STRUCTURE AND SENTENCE PROCESSING

Lindamood, & Hillyard, 1984; Schwanenflug& LaCount, 1988); these effects can beserved even for words that do not form accable sentence completions (Kleiman, 1980;tas & Hillyard, 1984; Kutas et al., 1984).

The types of words facilitated by a contand the degree of facilitation for each seemvary with the nature of the context itself. Fexample, highly constraining contexts seemprovide greater facilitation of “best comptions” than do less constraining conte(Fischler & Bloom, 1979; McClelland &O’Regan, 1981). At the same time, howevhighly constraining contexts have a narro“scope” of facilitation that does not extendless predictable items (e.g., SchwanenflugeLaCount, 1988; Schwanenflugel & Shob1985). Less constraining contexts, on the ohand, facilitate a wider range of items and pvide enhanced facilitation for less predictaitems. Research has also shown thatgreater semantic-associative information (more words that are lexically associated wittarget) in a context, one observes greater fitation of contextually congruent words (Duet al., 1989) and more elaborative inferendrawing (McKoon & Ratcliff, 1989); this factohas not always been controlled for in otstudies looking at, for example, effects of ctext on contextually incongruent, semanticaassociated targets.

Taken together, this body of work suggethat sentence contexts facilitate, in a gramanner, the processing of a set of concand/or words. Moreover, the nature and streof the sentence context affects what items/ccepts are facilitated and to what extent. Hoever, it cannot be information in the sentecontext alone2 that determines what is or is nfacilitated, as at times facilitation has beenserved for contextually inappropriate (butmantically related) items but not observedunexpected but contextually congruent o

2 Of course, no sentential context effects are whollyependent of long-term memory, as context effects nearily derive from information stored in memory. Whenpeak of the influence of sentence context alone, we rehe kinds of sentence context effects that would be exp
ven if memory were unstructured.
---

o

,r

,r

-

h.,

l-

-

dtsh-

-

--

s

(Schwanenflugel & LaCount, 1988). Becalanguage comprehension crucially relies onformation stored in long-term memory, we hpothesized that the structured nature ofmemory is another significant—but relativunexplored—variable likely to be affecting howords are processed during reading.

The N400 and Long-Term Memory

The hypothesis that the organization ofmantic memory plays an integral role in demining how information in a sentence contwill affect word processing receives suppfrom the observation that N400 effects, whinsensitive to nonsemantic manipulationscontext (e.g., changes in the physical attribof words (Kutas & Hillyard, 1980a) or grammatical and morphological violations (KutasHillyard, 1983)) or deviations in nonlinguiststimuli (e.g., anomalous notes in melodies (Bson & Macar, 1987)), do seem to be sensitivlong-term memory processes. N400 comnents have been recorded during investigatof recognition memory for both words (NevilKutas, Chesney, & Schmidt, 1986; SmStapleton, & Halgren, 1986) and pictu(Friedman, 1990). Some studies have claito observe N400-like components during meory tasks involving stimuli that are not particlarly semantic in nature. For example, Stusal. (1986) reported an N400-like componwhose amplitude varied with the numberpictures to be remembered in a continuousognition–memory task. Chao et al. (1995) areported what they describe as an N400 effeenvironmental noise stimuli, but only duriconditions involving repetitions after long dlays. Both groups suggest that their findiimplicate the N400 component in seathrough long-term memory.

Studies into the neurobiological basis ofN400 effect also support a link between tcomponent and long-term memory procesMcCarthy et al. (1995; also Nobre & McCarth1995) recorded field potentials from intracranelectrodes implanted in humans undergotreatment for epilepsy as they read sentencecomprehension. Anomalous sentence end

s-

tod

were associated with large field potentials in the

es.weipputhafiediancediaticaemoraacthave-on

tel-

tharoe oor

emontyerlre

99al

su

asiteonoctoorthere

sch. Inadex

nt,”lf,

ive

on-aveat-by

n-g tose-thete-atoreat-ob-ver-to

.,F ry,1 b,1 re-s lses al-i area sucha l.,1 xts,t jecta ablep thei ef-f vea ers,A s,A

T

ysi-o of as y bei ofk uret thep bys ceso ber-s s ac om-p the


left and right anterior medial temporal lobThe authors suggested that these potentialsgenerated in anterior fusiform and parahpocampal gyri and perhaps the hippocamproper. Grunwald et al. (1995) also showedthe presence or absence and amplitude ofpotentials recorded from the left anterior metemporal lobe were correlated with performaon a delayed word recognition task. As metemporal lobe structures are considered crifor successful performance on declarative mory tasks (e.g., Squire, 1987), a medial templobe source for at least some of the N400tivity at the scalp lends credence to the ideathe N400 may index semantic memory involment as a word is integrated with previous ctext.

Insofar as context effects—and associaN400 effects—derive from perceivers’ knowedge about the world and the access ofknowledge from memory, on-line language pcessing should be influenced by the structurthat memory. The structure of semantic memmay be based on many factors, but it selikely that an important part of its organizaticould involve the kind of featural similaristructure that has been observed to undhuman categorization (and information repsentation in the brain; see, e.g., Tanaka, 1for a striking example from higher order visurepresentation). Categorization researchgests that many human categories aretaxo-nomic: items are grouped together on the bof shared perceptual and functional attribu(e.g., Kay, 1971; Rosch, Mervis, Gray, Johns& Boyes-Braem, 1976) and these groupingscur at multiple levels of generality, similarbiological taxonomies. In this scheme, categmembership is graded, determined by whe—and how many—attributes an item shawith other members of a category (e.g., Ro1975; Rosch & Mervis, 1975; Rosch, 1973)turn, there also seems to be a structured, grorganization of categories themselves. Forample, a particular item may be called “a pla“a flower,” or “a rose”—each a category itsealbeit with successively decreasing inclusness.

Consistent with a relationship between taxoN

re-stldl

ll-l-t

-

d

t-f

ys

ie-6,

g-

ss,-

yr

s,

ed-

-

nomic categories, long-term memory, and ctext effects, electrophysiological studies hshown that N400 amplitude is sensitive to cegory membership. For example, in studiesPolich (1985) and Harbin et al. (1984) voluteers were shown a series of words belongina particular taxonomic category. The wordries ended either with another member ofcategory or with a word from a different cagory. Both groups found that a final word thwas not a category member generated mN400 activity than did a category member. Cegory effects on the N400 have also beenserved during the performance of sentenceification tasks (in which individuals are askedjudge the truth of statements of the form “AnXis aY”—for example, “A robin is a bird”) (e.g

ischler, Bloom, Childers, Roucos, & Per983; Kounios, 1996; Kounios & Holcom992). In these tasks, an enlarged N400ponse is observed to the final word of fatatements such as “A carrot is a fruit.” Revengly, in some cases large N400 responseslso observed at the end of true statementss “A carrot is not a fruit” (e.g., Fischler et a983). In other words, at least in some conte

he categorical relationship between the subnd object actually seems to be a more reliredictor of the N400 response than the fit of

tem in the sentence context itself (thoughects of propositional truth on the N400 halso been reported (Fischler, Bloom, Childrroyo, & Perry, 1984; Fischler, Childerchariyapaopan, & Perry, 1985)).

he Present Study

Taken together, behavioral and electrophlogical evidence suggest that the impactentence context on a word’s processing manfluenced by and interact with the structurenowledge in long-term memory—a structhat is likely based, at least in part, onerceptual and functional similarity capturedemantic categories. In fact, the first influenf both semantic context and category memhip on lexical processing are manifest ahange in the amplitude of the same ERP conent—the N400. Therefore, we can use

- 400 to examine the extent to which long-term

in-ingheem

irk.erin

ena

g th(1)athu

itat

andithla-

reas dcifieya

eyofexlaemo-

aplala-form.g

woateedulu

oreay,

ath

red

onse-

thed in

aint,on-.ob-andlp

t in-forug-ex-

e-thetex-ndested,rkio-

ex-iola-

x-se-

thetheea-nceto

ola-larsestcificin-x-

).toat

ilar


memory structure interacts with contextualformation during on-line sentence processIn particular, in this study we examined whetreaders’ processing would be affected by mory structure even when that structure wasrelevant to the language comprehension tasaddition, we aimed at getting a better undstanding of the role of memory structurereading by comparing its influence when stence contexts are strong versus when theyweaker.

We addressed these issues by comparineffects of two types of contextual violations:those that come from the same semantic cgory as the contextually predicted item and tshare many features in common with(“within-category violations”) and (2) those thcome from different semantic categoriesthus share far fewer features in common wthe predicted item (“between-category viotions”). ERPs were recorded as volunteerspairs of sentences. Each sentence pair wasigned to create an expectation for a speexemplar of a specific category (e.g., “Thwanted to make the hotel look more liketropical resort. So along the driveway thplanted rows of . . . ”). The second sentencethe pair ended with either (1) the expectedemplar (“palms”), (2) an unexpected exempfrom the same category as the expected explar (“pines”), termed the within-category vilation, or (3) an unexpected exemplar fromdifferent category than the expected exem(“tulips”), termed the between-category viotion. It is important to note that exemplarsthe between-category violations were still mebers of a shared higher-level category (e“plants”) and therefore match the other titems on most general dimensions. Items rotroles across sentences such that they serveach type of ending once across the stimset, as the following example illustrates:

1. They wanted to make the hotel look mlike a tropical resort. So along the drivewthey planted rows ofpalms/pines/tulips.

2. The air smelled like a Christmas wre
and the ground was littered with needles. Th
.r--In-

-re

e

e-s

de-c

-r

-

r

-.,

dass

land in this part of the country was just covewith pines/palms/roses.

3. The gardener really impressed his wifeValentine’s Day. To surprise her, he hadcretly grown someroses/tulips/palms.

4. The tourist in Holland stared in awe atrows and rows of color. She wished she livea place where they grewtulips/roses/pines.

The sentence contexts varied in their constrthe degree to which they led to a strong (csistent) expectation for the best completion

Comparing the pattern of ERP resultstained when individuals read sentences withwithout violations of the two types should heunravel the importance of sentence contexformation and semantic memory structurelanguage comprehension. Previous work sgests that the best completions (i.e., thepected exemplars) will elicit a positivity btween 300 and 500 ms. By contrast,between-category violations, which are contually unexpected, difficult to integrate, ashare few features in common with the bcompletions, will likely elicit an N400 in thsame time window (e.g., Kutas & Hillyar1980b). What is unknown from previous wois how the response to the within-category vlations will compare with the response topected exemplars and between-category vtions.

Within-category violations are similar to epected exemplars in that they share manymantic features in common. Therefore, if (atlevel of processing indexed by the N400)system is sensitive only to a fairly general fture match between an item and a sentecontext, one might expect a similar amplitudeexpected exemplars and within-category vitions. A difference between expected exempand within-category violations would suggthat the system is sensitive to more specontextual information (the kind that allowsdividuals, off-line, to predict the expected eemplar but not the within-category violation

Alternatively, if the system is sensitivespecific contextual information—and thalone—one would expect an N400 of sim
eamplitude to both within- and between-category

te-othonffellere-ugby

atecoproananugcifihiperm

c-ononofforn-ke

ofrd

loz(2tedee

la-

e thee acs foi n-

beibleanyon-

omriesl ofatenti-half

de-gyhirt,

asicar,oulderened

f thethesis-e-

uip--

thew adl dinb , &P

1 oral.,

theose

es inilar

hape.meForandoesit isomt itsms.oag-

raged arems,teria;actu-

vel.ough


violations. Neither within- nor between-cagory violations are expected. Moreover, bare relatively implausible in their sentence ctexts. There is thus no reason for them to dibased on their fit to context alone. A smaN400 amplitude to the within- relative to btween-category violations would therefore sgest that long-term memory is structuredfeature similarity as reflected in semantic cgories and that, independent of sentencetext, this structure impacts on-line sentencecessing. If within-category violations elicitN400 of intermediate amplitude (greater thexpected but less than between), it would sgest that the system is sensitive both to specontextual information and to the relations(feature overlap) between concepts in long-tmemory.

Finally, if, in fact, long-term memory struture affects on-line language processing,can examine its interaction with sentence ctext information by comparing the impactmemory structure when sentential context inmation is strong (in highly constraining setences) with its impact when context is wea(as in less constraining sentences).

METHODS

Materials

Stimulus material consisted of 132 pairssentences, each ending with three target wo(1) the expected exemplar, the highest cprobability ending for a given sentence pair,the within-category violation, an unexpec(cloze probability, 0.05) exemplar from thsame taxonomic category as the expectedemplar3, and (3) the between-category viotion, an unexpected (cloze probability, 0.05)

xemplar from a different category thanxpected exemplar. The first sentence of eentence pair established the expectationtem and category4. In contrast, the second se

3 While these items came from the same category,ere generally not lexical associates (only 10/132 h

exical association greater than 0.1 according to the Eurgh Associative Thesaurus (Kiss, Armstrong, Milroyiper, 1973)).

4
Forty-two out of 132 of these first-sentence contexts
-r

-

-n--

-c

e-

-

r

s:e)

x-

hr

tence, when separated from the first, couldplausibly completed by any of the three posstargets. There were no lexical associates ofof the possible endings within the sentence ctaining the target word.

Target items were pictureable objects fr66 categories (two items from each). Categowere chosen to be those at the lowest leveinclusion for which the average undergradustudent could be expected to readily differeate several exemplars. For approximatelythe categories used, this level was basic astermined by Rosch et al. (1976) or by analo(e.g., tree, fish, bird, cat, dog, pants, shoes, slamp, and car were all determined to be blevel by Rosch et al. and flower, rodent, beboat, insect, dinosaur, cheese, bread, etc. wseem to be so as well). Other categories wbased at what Rosch et al. would have defias the next highest level (a superordinate obasic level) because it was unclear thataverage participant could clearly and contently differentiate below this level (e.g., vegtable (different types of carrots?), sports eqment (different types of bats?))5. Between

ya-

contained a word lexically associated at a level of 0.greater (Edinburgh Associative Thesaurus (Kiss et1973)) with the expected exemplar.

5 Rosch et al. (1976) used four criteria to determinebasic level: the basic level is the most inclusive level whmembers (1) possess significant numbers of attributcommon, (2) have similar motor programs, (3) have simshapes, and (4) can be identified from the average sHowever, these criteria can be difficult to apply to socategories and may yield conflicting results for others.example, “country” is a category for which individuals cname a large number of exemplars and for which therenot seem to be a lower level other than exemplar. Butdoubtful that the category “country” can be identified frits average shape or what it would mean to say thamembers are interacted with via similar motor prograWhile “book” may fulfill the definition for basic, it is alsthe case that the category “reading material” (books, mazines, newspapers) can likely be identified from its aveshape and that its members have similar shapes aninteracted with via similar motor programs. Other itesuch as “hammer” clearly seem to be basic by these crihowever, the average undergraduate probably cannotally make (at least verbal) differentiations below this leFor the purposes of this study, therefore, to have en
categories we were forced to go to the next highest level

erke

ionndn-

ulume

thctlyty,eaforer

lista

lisctee-

theeatheceenerrean

32sinesoenoeewie

wa. A

ndtu-a

tiontic-tedgh-anarsbe-ozeob-a-ola-

msn-of

herheirin-

emineRPen-ndzetheility.784nthe

hade of

n-t gle,h ints ith al thee ener-a ncya on-s

ble

is,venottain

t bye cted


category targets for each sentence pair wchosen from a related category that sharedfeatures (e.g., animacy, size, general functwith that from which the expected exemplar awithin-category violation were derived. Appedix A lists the categories and their pairings6.

Target items were rotated across the stimset such that each item appeared three tionce as each kind of ending. Thus, acrossexperiment target conditions were perfecontrolled for length, frequency, imageabiliand concreteness; context sentences incondition were also perfectly controlledlength and grammatical complexity. The expimental sentences were divided into threeof 132 sentences each. Sentence contextsitems were used only once per list; eachconsisted of 44 of each type of target (expeexemplars, within-category violations, btween-category violations). Within each list,three target conditions were matched for mword length and frequency. To balancenumber of plausible and implausible sentenread by each participant, 44 plausible filler stence pairs were added to each list. Stimuli wrandomized once within each list and then psented in the same order for each participAppendix B gives examples of the stimuli.

Cloze Procedure

Cloze probabilities were obtained for the 1sentence pair contexts (sentence pairs misthe final word of the second sentence). Thwere divided into two lists so that the twsentence contexts presumed to be predictivitems coming from the same category didboth appear on the same list. Student voluntwere asked to complete each sentence pair“the first word that comes to mind.” List onwas completed by 56 students and list twocompleted by a different set of 59 students

(carpentry tools) in order to have clearly differentiaitems.

6 Items of a category vary in their “typicality”—thatthe degree to which they are judged to be representatithe category to which they belong. Typicality wasmanipulated in this study, and the set of items used con
both typical and atypical exemplars. (
ey)

ss,e

ch

-sndtd

n

s-e-t.

ge

oftrsth

s

subset of the original stimuli was rewritten aclozed separately by a third group of 55 sdents. Cloze probability for a given word ingiven context was calculated as the proporof individuals choosing to complete that parular context with that particular word. Expecexemplars were always the item with the hiest cloze probability for a given context. Mecloze probability for the expected exemplwas 0.74. Within category violations andtween category violations always had clprobabilities of less than 0.05. Mean cloze prability was 0.004 for the within-category violtions and 0.001 for the between-category vitions.

Constraint

Although all expected exemplars were itewith the highest cloze probability for their setence contexts, the actual cloze probabilitythese items ranged from 0.17 to 1.0. In otwords, the sentence contexts differed in tconstraint, or the degree to which they leddividuals to strongly expect one particular itversus a number of different items. To examthe effects of sentential constraint on the Eresponse to target items, we divided the stences into two groups, “high constraint” a“low constraint,” by a median split on the cloprobability of the expected exemplar. Forhigh constraint sentences, the cloze probabof the expected exemplars had a range of 0to 1.0 and an average value of 0.896 (media50.904). For the low constraint sentences,cloze probability of the expected exemplarsa range of 0.17 to 0.784 and an average valu0.588 (median5 0.608). High constraint seences are thus those in which there is a sinighly preferred ending, while low constraentences are those that are compatible warger range of ending types and in whichxpected exemplar has at least one, and glly several, close competitors. Word frequend word length were controlled across all ctraint and ending type conditions7.

of

s

7 Average values for word frequency/word length splinding type and constraint were: high constraint, expe
18.6/6.0); low constraint, expected (21.2/6.0); high con-

owesd avelaor

enratirncthonereenc

0%atenusm-enici-or

byofndantilitade-

ibilresarionm

en-ef-

. exp ntlm ns[ -

gory

f-are

a-intpela-ledintg

Ts ac-t

tedon-th

w

. Ingsral8;

orete-

gra-

ersge,

t fort inht-ntoryk-

orthe

2);ee

rom

t


Plausibility Ratings

The two violation types were both kept bela cloze probability of 0.05. While this indicatthat none of the violations was consideregood ending for the sentence context, it leaopen the question of whether one of the viotion types might have been, on average, a mplausible ending. To determine this, a differgroup of student volunteers was asked tothe plausibility of all of the endings within theexperimental sentence contexts. The sentewere split into the same three lists used inactual ERP experiment so that no item or ctext was repeated within a list. Volunteers wasked to rate how much “sense” each sentpair made on a percentage scale (wheremeant the sentence pair “makes no sense(is very implausible)” and 100% meant the stence pair “makes perfect sense (is very plable)”). Lists one, two, and three were copleted, respectively, by 18, 21, and 18 studvolunteers; none of these individuals partpated in either the cloze probability ratingsthe ERP experiment.

Mean rated plausibility was calculatedaveraging the plausibility ratings for all itemsa given condition within each participant athen averaging the scores across participExpected exemplars had a mean plausibrating of 95.6%, within-category violations ha mean plausibility rating of 28.3%, and btween-category violations had a mean plausity rating of 15.3%. These plausibility measuwere subjected to an omnibus analysis of vance (ANOVA) with repeated measuresthree levels of Ending Type (expected exeplars vs within-category violations vs betwecategory violations), revealing a significantfect of Ending Type [(F(2,112)5 2738.25;p ,001]. Planned comparisons showed thatected exemplars were rated as significaore plausible than within-category violatio

t 5 46.06;p , .001] and within-category vio

straint, within (21.0/5.9); low constraint, within (18.8/6.high constraint, between (18.7/6.0); low constraint, betw(21.1, 6.0). Word frequency information was obtained f
Francis and Kucera (1982).
s-ete

ese-

e

all-i-

t

s.y

-

-

-

-y

lations as more plausible than between-cateviolations [t 5 15.75;p , .001].

Plausibility ratings for the violation types ater splitting the sentences by constraintshown in Table 1.

An omnibus ANOVA with repeated mesures performed on two levels of Constra(high vs low) and three levels of Ending Ty(expected exemplars vs within-category viotions vs between-category violations) reveasignificant main effects of both Constra[F(1,56) 5 3472.05; p , .001] and Endin

ype [F(2,112) 5 1369.32;p , .001] and aignificant Constraint by Ending Type interion [F(2,112) 5 994.49; p , .001]. Ratedplausibility significantly increased for expecexemplars in high versus low constraint ctexts [t 5 5.00;p , .001] but decreased for bo

ithin-category violations [t 5 3.54;p , .001]and between-category violations [t 5 8.21;p ,.001] in high versus low constraint contextsother words, the pattern of plausibility ratinwas congruent with claims from the behavioliterature (Schwanenflugel & LaCount, 198Schwanenflugel & Shoben, 1985) that mhighly constraining contexts allow greater ingration of best completions but reduced intetion of improbable completions.

Participants

Eighteen UCSD undergraduate volunte(10 men and 8 women, 18 to 24 years of amean age 20) participated in the experimencredit and/or cash (none of these took parany of the norming procedures). All were righanded (as assessed by the Edinburgh Inve(Oldfield, 1971)), monolingual English speaers with no history of reading difficultiesneurological/psychiatric disorders; five ofn

TABLE 1

Highconstraint

Lowconstrain

Expected exemplars 97.7% 93.5%Within-category violations 23.6% 30.2%Between-category violations 11.9% 18.7%

volunteers reported having a left-handed family

as

enctrinntlusin

heyvehedeionntaith

nge athcoestheenen

orynereeresrge00a

ortthntist, 210

witithfiedeeol-gectensen

re-an

oid.sitesdftf),),dd

htnde),t-

ghtc-lale

deserv-achode

at ainu-ard

ge-alssivee-elyts.o-m-

foreset.cu-tede-of


member. Six participants were randomlysigned to each of the three stimulus lists.

Experimental Procedure

Volunteers were tested in a single experimtal session conducted in a soundproof, elecally shielded chamber. They were seatedcomfortable chair approximately 60 cm in froof a monitor and instructed to read the stimusentences for comprehension. They wereformed at the start of the experiment that twould be given a recognition memory test othe stimuli at the conclusion of recording. Tsession began with a short practice trialsigned to reiterate the experimental instructand to acclimate volunteers to the experimeconditions and the task. Each trial began wthe first sentence of a sentence pair appearifull on a CRT. Volunteers read this sentenctheir own pace and pushed a button to viewsecond sentence. Presentation of the sesentence was preceded by a series of crossorient the volunteer toward the center ofscreen. The second sentence was then presone word at a time in the center of the screNonsentence final words were presented fduration of 200 ms with a stimulus-onset aschrony of 500 ms. Sentence final words wpresented for a duration of 500 ms. Voluntewere asked not to blink or move their eyduring the second sentence. The final, taword was followed by a blank screen for 30ms, after which the next sentence appearedtomatically. Volunteers were given a shbreak after every 17 pairs of sentences. Atconclusion of the recording session, participawere given a recognition memory test consing of 50 sets of sentence pairs: 10 new onesunchanged experimental pairs (of whichended with expected exemplars, 5 endedwithin-category violations, and 5 ended wbetween-category violations), and 20 modisentence pairs in which the final word had bchanged from that originally viewed by the vunteer (10 in which violations had been chanto expected exemplars and 10 in which expeexemplars had been changed to violatioVolunteers were instructed to classify the s
tences as new, old, or similar (changed).
-

-i-a

-

r

-sl

intendto

ted.a-

s

t

u-

es-0

h

n

dd).-

EEG Recording Parameters

The electroencephalogram (EEG) wascorded from 26 tin electrodes embedded inElectro-cap, referenced to the left mastElectrode sites are shown on Fig. 1. Theseincluded midline prefrontal (MiPf), left anright medial prefrontal (LMPf and RMPf), leand right lateral prefrontal (LLPf and RLPleft and right medial frontal (LMFr and RMFrleft and right mediolateral frontal (LDFr anRDFr), left and right lateral frontal (LLFr anRLFr), midline central (MiCe), left and rigmedial central (LMCe and RMCe), left aright mediolateral central (LDCe and RDCmidline parietal (MiPa), left and right mediolaeral parietal (LDPa and RDPa), left and rilateral temporal (LLTe and RLTe), midline ocipital (MiOc), left and right medial occipita(LMOc and RMOc), and left and right lateroccipital (LLOc and RLOc). Blinks and eymovements were monitored via electroplaced on the outer canthus (left electrode sing as reference) and infraorbital ridge of eeye (referenced to the left mastoid). Electrimpedances were kept below 5 KV. EEG wasprocessed through Grass amplifiers setbandpass of 0.01–100 Hz. EEG was contously digitized at 250 Hz and stored on hdisk for later analysis.

Data Analysis

Data was rereferenced off-line to the albraic sum of the left and right mastoids. Tricontaminated by eye movements, excesmuscle activity, or amplifier blocking were rjected off-line before averaging; approximat10% of trials were lost due to such artifacBlinks were corrected via a spatial filter algrithm devised by Dale (1994). ERPs were coputed for epochs extending from 100 ms bestimulus onset to 920 ms after stimulus onAverages of artifact-free ERP trials were callated for each type of target word (expecexemplars, within-category violations, btween-category violations) after subtraction
the 100 ms prestimulus baseline.

fiedtheon

m-alllyd

esre

esersororm-es”

par-hatded

lun-rd-o-iormsmsmsingingare

on,tralion

ed,ntr


RESULTS

Behavior

On average, volunteers correctly classi88% (range 74 –100%) of the items onrecognition memory test. The most commtype of error was a misclassification of “siilar” sentences (those in which only the finword had been altered from that actuashown in the experiment) as “old,” followeby a misclassification of “old” sentenc(those seen in the same form during thecording session) as “similar.” Together, thtwo error types account for 73% of all erroobserved. Most of the remainder of the errconsisted in volunteers classifying “old”“similar” sentences as “new” (average nuber of either of these type of errors was lthan one). Only two errors in which “new

FIG. 1. Schematic of the electrode array usedarranged in a series of four equally spaced conce

sentences were classified as “old” or “similar”

-

s

s

(one each) were observed across the 18ticipants. The behavioral results indicate tthe experimental sentences were attenduring the recording session.

ERPs

Grand average ERPs (across all 18 voteers) to sentence final words from all recoing sites are shown in Fig. 2. Early compnents in all conditions include, at postersites, a positivity peaking around 110(P1), a negativity peaking around 180(N1), and a positivity peaking around 280(P2), and, at frontal sites, a negativity peakaround 130 ms (N1) and a positivity peakaround 230 ms (P2). Early componentsfollowed, in the expected exemplar conditiby a broad late positivity, largest over cenand posterior sites, and, in the two violat

e experiment. In all, 26 scalp electrodes were employic rings.

in th

conditions, by a negativity peaking around

anonte-

owrtain

.0- toryas

that


400 ms (N400), also largest over centralparietal sites. The N400 in the two violaticonditions is followed by an extended lapositivity of similar amplitude to that ob

FIG. 2. Grand average (N 5 18) ERP waveformsNegative is plotted up. The ending types are char450-ms time window, expected exemplars (solid lviolations (dashed line) and between-category violarger for the between category violations. A box awill be used in subsequent figures.

served for the expected exemplars.

dPeak Latency of the N400 Response

In order to determine the appropriate windfor mean amplitude analyses and to asce

the three ending types shown at all 26 electrode siteserized by the same set of early components. In the 35showed a sustained positivity while both within-categons (dotted line) showed a negativity, the N400, which wnd the right medial central site indicates the electrode

foractine)latiorou

that the latency of the N400 did not differ across

eaea

adeemen-deseenwi

85al

n),75ecnt

ct

eawinostct

aste

ypla-ev-

cttedetivexeateand bce

vi-re2,

ttlyals.ndof

indidssedto-

-

iin

vedos-g-vi-esriortlyere

viag

in-la-

oferete-

a oryv ane. ac-t. hee istri-b in-c

ms( fectc nseta ubseto ingc as thec asice ante 50–


conditions, latency of the largest negative pbetween 350 and 450 ms was measured forcondition in each participant and subjected toomnibus ANOVA. Repeated measures incluthree levels of Ending Types (expected explars vs within-category violations vs betwecategory violations) and 26 levels of ElectroAll p-values in this and all subsequent analyare reported after Epsilon correction (Grehouse–Geisser) for repeated measuresgreater than one degree of freedom.

Mean peak latency (in milliseconds) is 3for the expected exemplars (though, in factmost no N400 is observed in this conditio377 for the within category violations, and 3for the between category violations. The effof Ending Type was not statistically significa[F(2,34)5 2.15;p 5 .157] and did not interawith the effect of electrode [F(50,850)5 0.92;p 5 .511].

Mean Amplitude Analyses

Based on the peak latency analysis, mvoltage measures were taken in a 50-msdow around 375 ms (i.e., 350–400 ms pstimulus-onset). These measures were subjeto an omnibus ANOVA. List (three levels) wa between-subjects variable while repeameasures included three levels of Ending T(expected exemplar vs within-category viotion vs between-category violation) and 26 lels of Electrode.

Effects of list.While there was no main effeof List [F(2,15)5 0.67;p 5 .524], a significanList 3 Ending Type interaction was observ[F(4,30)5 4.59;p 5 .005]. Examination of thmeans revealed no difference in the qualitapattern of ending type effects; in all casespected exemplars have the most positive mvoltage in this time window and between-cagory violations have the least positive mevoltage. The interaction seemed to be causedifferences in the range of these differenwhich are largest in list 3 (7.26, 3.66, and20.22mV for expected exemplars, within-categoryolations, and between-category violations,spectively), and smallest in list 2 (list 1: 5.53.10, 0.88mV; list 2: 3.48, 1.44, 0.80mV). List
was also found to interact with Electrode5
kchnd-

.s-th

-

t

n--ed

de

e-n

-

y,

-

[F(50,375)5 4.14, p 5 .002], indicating thathe distribution of the N400 effect is slighdifferent across lists and thus across individuGiven that individuals’ brainwaves vary athat there were a relatively small numberparticipants per list (n 5 6), these variationsthe overall size and distribution of effectsnot seem important for the questions addreby this study; lists were therefore groupedgether for subsequent analyses.

Effects of ending type.A main effect of Ending Type was observed [F(2,30)5 55.03,p ,.001], as was an Ending Type3 Electrodenteraction [F(50,750)5 7.04,p 5 .001]8. Fig-ure 3 shows this effect of Ending Type,which the smallest N400 amplitude is obserto the expected exemplars, which are very pitive in this time window, and the largest neativity is observed to the between-categoryolations. The N400 to the two violation typtended to be largest over central and postesites, larger medially than laterally, and slighlarger over the right than the left hemisphsites. Planned comparisons were conductedan omnibus ANOVA on two levels of EndinType (between-category violation vs withcategory violation and within-category viotion vs expected exemplar) and 26 levelsElectrode. Between-category violations wsignificantly more negative than within-cagory violations [F(1,17) 5 17.06, p 5 .001]

cross the scalp. Additionally, within-categiolations were significantly more negative thxpected exemplars [F(1,17) 5 32.75, p ,

001]; this effect showed a significant interion with Electrode [F(25,45) 5 6.50, p 5004], indicating that the broad positivity to txpected exemplars has a different scalp dution than the N400 response to the withategory items.

8 Analyses were done in a 50-ms window around 375the peak of the N400 effect). The beginning of the efould be observed from about 250 ms post-stimulus-ond ended about 250 ms later. We chose to analyze a sf that time interval to minimize the effects of overlappomponents and because some of the effects (suchonstraint effects) were more temporally specific. The bnding type effect, however, was statistically significven when analyzed over large time windows (e.g., 2
00 ms: [F(2,34)5 35.16;p , .001]).

chtheththeceusgoP

urendurypvssior).edin

e

. rd amp esto verc itesa thel n[ 0e cci-p ig-g byL

l andleftusites

leftthe

pi-byen,

in

u-to

re-tivedis-

ob-rms.dial,wasd intrala).

anon

eevs

ory

int

ialc amp nific ine)t nif-i x-p


Distribution of the N400 Effect

The distribution of the N400 effect for eaviolation type was examined by looking atmean amplitude ERP difference betweenviolation type and the expected exemplar in350- to 400-ms time window. The differenwaves (within-category violation ERP minexpected exemplar ERP and between-cateviolation ERP minus expected exemplar ERwere normalized according to the proceddescribed in McCarthy and Wood (1985) awere then subjected to an ANOVA on forepeated measures, two levels of Ending TDifference, two levels of Hemisphere (leftright), two levels of Laterality (lateral vmedial), and four levels of Anterior/Poster(prefrontal vs frontal vs parietal vs occipital

A main effect of Laterality was observ[F(1,17) 5 34.57; p , .001] as was a ma

FIG. 3. Effect of ending type, shown at the right medentral site. A three-way split can be observed in thelitude of the N400 response. N400 amplitude was sigantly larger for between-category violations (dotted lhan for within-category violations (dashed line) and sigcantly larger for within-category violations than for eected exemplars (solid line).

ffect of Anteriority [F(3,51) 5 5.08; p 5

e

ry)

e

036]; there was a nonsignificant trend towaain effect of Hemisphere [F(1,17) 5 3.87;5 .066]. In short, N400 effects were bigg

ver medial relative to lateral sites and oentral/posterior relative to more anterior snd tended to be bigger on the right than on

eft. A Laterality by Anteriority interactioF(3,51)5 11.13;p , .001] indicates that N40ffects at lateral sites are biggest over the out while N400 effects at medial sites are best parietally. A trend toward a Hemisphereaterality effect [F(1,17) 5 3.61; p 5 .075]

suggests that the difference between medialateral electrode sites is greater over thescalp than over the right. N400 effects thtended to be largest over medial, parietal sand bigger over the right than over thehemisphere. This distribution can be seen inERPs in Fig. 4; it is the one that is most tycally reported for N400 effects during wordword sentential reading (Kutas & Van Pett1994).

A main effect of Ending Type was agaobserved [F(1,17) 5 4.52; p 5 .048]. EndingType did not interact with any of the distribtional factors. Thus, the N400 responsewithin-category violations and the N400sponse to between-category violations relato expected exemplars are very similar intribution.

Mean Amplitude Analyses of Constraint

Effects of contextual constraint could beserved in the grand average ERP wavefoThese effects were most prominent over mecentral-parietal sites where the N400 effectbiggest. Therefore, constraint was analyzethe same time window at the four medio-cenelectrode sites (MiCe, LMCe, RMCe, MiPMean voltage measures were subjected toomnibus ANOVA with repeated measurestwo levels of Constraint (high vs low), thrlevels of Ending Type (expected exemplarswithin-category violations vs between-categviolations), and four levels of Electrode.

Main effects of both Ending Type [F(2,34)538.43; p , .001] and Constraint [F(1,17) 56.47;p 5 .021] were observed. High constra

--

sentences are associated with overall higher

lows

onin

on-

beow

gerall

nceenw

nnedand

type

-,

on


(more positive) mean amplitudes than areconstraint sentences. In addition, there waConstraint by Ending Type interacti[F(2,34) 5 3.45; p 5 .043], as can be seenFig. 5. Examination of the means (high cstraint: 8.43, 5.10, and 1.23mV for expectedexemplars, within-category violations, andtween-category violations, respectively; l

FIG. 4. Difference waves showing the N400 efcategory violations (dotted lines). The waveformsLDFr, LDPa, LMOc, RMPf, RDFr, RDPa, RMOc,N400 effect. For both conditions, the N400 effectthe right than on the left.

constraint: 7.72, 2.38, and 0.85mV) revealed

a

-

that, while mean amplitudes are slightly larfor high than low constraint sentences forending types, most of the amplitude differeis accounted for by the difference betwewithin-category violations in high versus loconstraint sentences as shown in Fig. 6. Placomparisons were performed between highlow constraint sentences for each ending

t to within-category violations (solid line) and betweenthe 16 electrode sites (LLPf, LLFr, LLTe, LLOc, LMPff, RLFr, RLTe, RLOc) illustrate the distribution of the

s larger over medial posterior sites and slightly larger

fecat

RLPwa

using t-tests (df 5 17; a 5 0.05). These con-

x-ainn-

c n-

s hm te-g ent lowc

S

ine os-s ner-a t isp theo overi ovep ifi-c ef-f derb er-e idea t oft tingo di-t la-t ec-o gem in a

otin-an

ces

a-00intnce


firmed that the difference in the N400 to epected exemplars in high versus low constrsentences (t 5 0.91;p 5 .373) and to betweeategory violations in high versus low co

FIG. 5. Effect of constraint on the N400 responaffect the response to expected exemplars (solicategory violations (dashed line) in high constrawithin-category violations in low constraint senten

FIG. 6. Effect of constraint on within-category violtions, shown at the right medial central site. Larger N4were elicited by within-category violations in low constrasentences (dotted line) than in high constraint sente

5(solid line).

ttraint sentences (t 5 0.50;p 5 .625) are botinimal. By contrast, the N400 to within-caory violations is significantly smaller wh

hese are in high constraint rather thanonstraint sentences (t 5 3.91;p 5 .001).

tability of Effects over Items

Because of the low signal to noise ratioslectrophysiological data, it is usually not pible to perform item analyses, as would gelly be done for behavioral data. However, iossible to provide some indications thatbserved effects are reasonably stable

tems. The list and constraint analyses abrovide some evidence for stability, as signant and qualitatively similar ending typeects were observed for all three lists and unoth constraint conditions (consisting of diffnt contexts and different items). To provdditional evidence, we made another spli

he data for each participant into bins consisf the first 22 items of each ending type con

ion (expected exemplar, within-category vioion, between-category violation) and the snd 22 items of each condition. Mean voltaeasures were then taken for each bin

shown at the right medial central site. Constraint did nne) or between-category violations (dotted line). Withsentences (left) elicited smaller amplitude N400s th(right).

s

s

se,d liint

0-ms window around 375 ms (i.e., 350–400

wetedal

ingvi-26

ing-

iin

ee

t.T bet the( ntlya mp app itht ly-s abs that

si-t thw teg rN trar oft in-c in-c n-s

t tod g ot nlyb byp bua t inf tlyd l-e esi linep

n i.e.,b redw sis-t theo ofw ext( dif-f rity( bleo Wea ofe byt d ap int)v on-s

a-t bya toa vedt em-b s inc emi ex-e oryv latep one 600m aset rei oryw erys notw on-t ificf e,c cta-t

ce allys lsot m iss eent er-e otho ndt anys and


ms post-stimulus-onset). These measuressubjected to an omnibus ANOVA with repeameasures on two levels of Experimental H(first half vs second half), three levels of EndType (expected exemplar vs within-categoryolation vs between-category violation), andlevels of Electrode.

We again observed a main effect of EndType [F(2,34)5 33.65;p , .001] and an Endng Type3 Electrode interaction [F(50,850)57.61; p , .001]. However, neither the ma

ffect of Experimental Half [F(1,17) 5 2.82;p 5 .111] nor its interaction with Ending Typ[F(2,34) 5 0.48; p 5 .625] was significan

hus, there was no significant differenceween the ending type effect observed forrandom) selection of items that consisteppeared in the first half of each list as coared with that observed for the items thateared in the second half. This, combined w

he findings from the list and constraint anaes, suggests that our effects are reasontable over items and suggests in additionhe effects are stable over time/practice.

Summary of Main Results

While expected exemplars elicit a late poivity in the 350- to 400-ms time window, boithin-category violations and between-caory violations elicit a qualitatively simila400 response with a medial, posterior-cen

ight hemisphere distribution. The amplitudehe N400 is bigger for between- than for withategory violations and is bigger for withategory violations in low than in high cotraint sentences.

DISCUSSION

At a general level, this experiment soughetermine the extent to which the processin

he final word in a sentence is affected not oy specific information directly activatedrior words in the sentence (i.e., context)lso by more general, context-independen

ormation (semantic feature overlap) indireceriving from the structure of real-world knowdge in long-term memory. We addressed th

ssues by examining (1) whether the on-
rocessing of two items sharing significanto
re

f

-

--

lyt

-

l,

f

t-

e

umbers of semantic features in common (oth members of the same category) diffehen the preceding context was more con

ent with one than the other, and (2) whethern-line processing of two items, neitherhich is especially consistent with the cont

i.e., contextually expected), nonethelessered as a function of their semantic similacategorical relationship) to the most probar expected ending (i.e., best completion).lso examined whether or not the impactither of these variables would be modulated

he degree to which the context anticipatearticular exemplar (high contextual constraersus several possibilities (low contextual ctraint).Our results show that fairly specific inform

ion from the sentential context is availablet least 375 ms after a word’s presentationffect its processing. Specifically, we obser

hat the brain’s response to two category mers (items sharing many semantic featureommon) does differ inasmuch as one of ths consistent with the context (the expectedmplar) and the other is not (the within-categiolation). The expected exemplar elicited aositivity whereas the within-category violatilicited a moderate N400 between 300 ands. If processing the context serves to incre

he availability of only fairly general featunformation, then two members of a categhich share these features should elicit a vimilar brain response. However, this washat we observed, indicating instead that c

ext serves to increase the availability of speceature information; that is, as we will arguontext leads to specific predictions or expeions.

While contextual information specifinough to distinguish between two semanticimilar items affects word processing, it is ahe case that the language processing systeensitive to the categorical relationship betwhem. Specifically, there is a significant diffnce in the brain’s response to two words, bf which are inconsistent with the context a

hus unexpected, but one of which shares memantic features with the expected ending
ne of which does not. The N400 to an unex-

caor

00rend

ernra

tedbeenn-thetror-lsofhiith

als).o

nofaem

eficteexhisnlifict aIn

proat

on-estoth

uslnebeteplesinxtupi

on-

t inac-

menfor-ate-ere

.g.,if

utashe

ob-uldto-

aredforenot

wereas

hinlarsreduntm-is

nti-orer-ated

in-ate-nticri-ughom-n-tednticted

eto

te-to

wesys-

hiphin-the


pected word that is a member of the sameegory as the expected ending (within-categviolation) is significantly smaller than the N4to a word that is similarly unexpected but shafewer semantic features with the expected eing (between-category violation). This pattof results is in accord with reports of behaviofacilitation for items semantically associawith the most expected endings or so-calledcompletions (e.g., Kleiman, 1980; Schwanflugel & LaCount, 1988). However, since cotext and semantic similarity both modulatesame ERP component, the N400, the elecphysiological data allow two additional infeences to be drawn. First, this finding reveasignificant temporal overlap in the influenceboth these factors on a word’s processing. Thas been previously shown, although not was much care for the stimulus materials (seeKutas & Hillyard, 1984; Kutas et al., 1984Second, we can infer that the influencescontext and semantic feature overlap oword’s processing are of the same kind. Insas the effects of these variables differ, it seto be a matter of degree.

What is the basis for the processing bengained by items that are related to the expeexemplars, but which themselves are notpected? A straightforward explanation for twould have emerged if it were the case that ocategory-level (as opposed to more speccontextual information was available to affecword’s processing in the N400 time window.that case, we would expect an equivalentcessing benefit to accrue to members of a cgory regardless of their specific fit to the ctext; in this experiment, we would expect bcompletions and within category violationselicit similar ERPs. In fact, as in this study, boreaction time and ERP studies have previoobserved that unexpected, related items geally yield a response intermediate in valuetween a best completion and an unexpecunrelated ending. In other words, best comtions have generally enjoyed greater procesbenefits than semantically related but conteally unexpected endings, although both tycally show facilitated processing relative to c
textually unexpected and semantically unrelate
t-y

s-

l

st-

-

a

s

o

fars

td-

y)

-e-

t

yr--d,-g-

-

endings. It is important to note, however, thathese previous studies “related” items weretually related in several different ways. Sowere related because they shared feature imation (i.e., were in the same taxonomic cgory), as in the present study, but others wonly associatively or thematically related (e“umbrella” and “rain”) and thus shared few,any, semantic features in common (e.g., K& Hillyard, 1984; Kutas et al., 1984). Under tcircumstances, the intermediate responseserved for semantically related endings cohave been a spurious artifact of averaginggether a subset of related items that shgeneral features with the context (and therewere facilitated) and another subset that didshare any semantic features (and thereforenot facilitated). In our experiment, however,the nature of the relation between the witcategory violations and the expected exempwas controlled such that they always shamany semantic features, this particular accodoes not apply. Our results thus not only deonstrate that specific contextual informationavailable to distinguish between such semacally similar items in real time but also call fan alternative, viable explanation for the intmediate response to unexpected but relitems.

We believe that the smaller N400 to withcategory violations compared to between-cgory violations reflects an influence of semamemory structure, built of real-world expeence, on on-line language processing. Althothese two ending types are both equally incpatible with the specific information in the cotext that leads to the prediction for the expecexemplar, one of them has substantial semafeature overlap with the ending most predicin the context while the other does not. Wsuggest that it is the processor’s sensitivitythis featural overlap that affords within-cagory violations a processing benefit relativethe between-category violations. Moreover,are impressed by the sentence processingtem’s sensitivity to this categorical relationsbetween the expected ending and the witcategory item even though it is irrelevant to
dprocessing of the particular context and thus to

tha

harop

erplaesanltsa

ctete

edtetheehe

n al fewl tot hus fivw t tos ef-f n as l-l &M net llyc thee hlyu tede ngM eren oryv anyd icaa acc te-g oryv

b-s iont psw erN bec

be-rlyudeof),en-ort,asnotity;riesto

(asdlyzer-urtheinthee-e.rn00to

deand

everthee andtherat-rg-gs,atehisdearel-

tlyowiatedtedio-es,

ostur


making sense of the sentence (e.g., the factpines and palms are categorically relatedtrees is in no way relevant for determining tpalms are an appropriate adornment for a tical resort).

Before making this case strongly, howevwe need to consider possible alternative exnations for the N400 difference between thtwo violation types. As it happens, we creadily rule out the possibility that our resusimply reflect lexical associative priming fromword in the sentence context to the expeexemplar and, by extension, to the within-cagory violation (via some form of mediatpriming). Our sentence pairs were construcso that the sentence context ending withtarget was equally compatible with all thrending types (e.g., “He asked his friend ifcould borrow themagazine/book/pencil”), so

one of these target sentences containedexical associate to any ending type. Theexical associates that did exist were limitedhe first sentence of the pair and were teparated from the target word by at leastords or more. This distance is too greaustain typical lexical associative primingects, which are known to dissipate with eveingle intervening word (Gough, Alford, & Hoey-Wilcox, 1981; Masson, 1991; Ratcliff

cKoon, 1988). Furthermore, only about ohird of our context (first) sentences actuaontained a word lexically associated withxpected ending. Thus, we consider it hignlikely that the lack of an N400 to the expecndings reflects lexical associative primiore importantly, our expected exemplars wot lexically associated with the within categiolations overall (less than 10% containedegree of association). Thus (mediated) lexssociative priming cannot be invoked toount for the smaller N400s to the within-caory violations relative to the between-categiolations.Another, more likely, explanation for the o

erved N400 difference between our violatypes might be in terms of plausibility. Perhaithin category violations elicited small400s than between category violations
ause they were more plausible, i.e., actuall
atst-

,-

e

d-

de

ny

se

.

l-

-

did fit the context better. The relationshiptween N400 amplitude and plausibility is poounderstood. On the one hand, N400 amplithas been found to vary with the predictabilitya word in its context (Kutas & Hillyard, 1984which can arise from discourse as well as stence level factors (e.g., Van Berkum, Hago& Brown, in press). On the other hand, it hlong been known that N400 amplitude canserve as a pure index of global level plausibilFischler et al. (1983) showed in an early seof studies that the N400 is not sensitivenegation. As plausibility and expectancymeasured by cloze probability) are undoubtelinked, and as the N400’s sensitivity to cloprobability is well established, plausibility cetainly must make some contribution to oN400 effects. The real question, however, isextent to which plausibility alone can explathe observed reduction in N400 amplitude towithin-category violations. As will soon bcome evident, plausibility alone will not suffic

If plausibility alone were driving the patteof results observed, then at minimum N4amplitudes should be monotonically relatedrated plausibility. That is, N400 amplitushould decrease when plausibility increasesincrease when plausibility decreases, whatthe exact relation in the rate of change oftwo. At the most general level, we do observmonotonic relation between plausibility aN400 amplitude: best completions elicitsmallest N400s and the highest plausibilityings, between-category violations elicit the laest N400s and the lowest plausibility ratinand within-category violations are intermedion both variables. However, we find that tmonotonic relation between N400 amplituand plausibility does not hold when the databroken down by contextual constraint. Athough the rated plausibility is significanhigher for best completions in high versus lconstraint sentence contexts, the assocERPs do not differ. Likewise, although the raplausibility is higher for between category vlations in low than high constraint sentenctheir N400s also do not differ. Finally, and mdamning for the plausibility hypothesis, is o
yfinding that while the rated plausibility for

tlyn-

er-orew

argor

epiteudebil-of

d-ongengn-allhad

ealhisecfo

si-etioto

d-alth

ry.to

sstedith

thet uexla-hesndnk

w

or-actthisg alvedBys-

herereof

Infea-arilytethe

ord.ect-

bytheiondingthe

em-ent

anyh-

hist of

rentbutAssys-is)

rrentThelic-

hatins

mis-pre-is aourla-nt.of

n-y of


within category violations is also significanhigher under low than high contextual costraint, N400 amplitudes are significantly diffent in the opposite direction. That is, the mplausible within-category violations in the loconstraint sentences are associated with a lN400 than are the less plausible within-categviolations in the high constraint sentences.

If we loosen the montonicity criterion, wcan maintain the plausibility hypothesis desthe absence of a significant N400 ampliteffect in the presence of a significant plausiity effect. However, no explanation in termsplausibility will be able to account for our fining that among the within-category violatiendings, the more plausible endings elicit larN400s than do the more implausible endi(e.g., “baseball,” in “‘Checkmate!’ Rosaline anounced with glee. She was getting to be regood at baseball.” has a smaller N400 t“earring,” in “She keeps twirling it around anaround under her collar. Stephanie seems rhappy that Dan gave her that earring.”). Tdeviance from monotonicity forces us to rejthe plausibility hypothesis as an explanationthe current pattern of results.

While lexical associative priming and plaubility clearly play significant roles in on-linlanguage comprehension, even in combinathey cannot account for the smaller N400swithin- than between-category violation enings. We now return to consider our originhypothesis that the explanation is inherent instructure of information in long-term memoThe greater reduction in N400 amplitudewithin category violations in more than leconstraining contexts is counter to their raplausibilities. Instead, it appears to pattern wthe expectancy and plausibility ratings forexpected items: the very contexts that severy specific expectations for the expectedemplar also provide the within-category viotions with the greatest facilitation (i.e., tgreatest N400 reduction). Thus, there seembe a functional link between the expected eings and the within-category violations—a liwhich we argue reflects memory structure.

This functional link has implications for ho
information in a sentence context is used durin
ery

rs

yn

ly

tr

n

e

p-

to-

language comprehension as well as for theganization of long-term memory and its impon sentence processing in real time. First,link suggests that in the course of processinsentence, the comprehension system is invoin some process tantamount to prediction.using the word “prediction,” we do not necesarily mean to imply that the process is eitconscious or strategic. Rather, prediction hrefers to activation of the semantic featuresupcoming words prior to their occurrence.this specific case, we mean that semantictures of the category exemplar (not necessa specific lexical item) most likely to complethe target sentence are activated prior topresentation of the actual sentence-final wWhen the prediction is incorrect and the expancy is not met, the data are characterizedincreased N400 activity relative to whenprediction is correct. Note that such activatof the semantic features of the expected enoccurs above and beyond the activation ofsemantic features of the context words thselves, although naturally it must be contingon their presence.

Of course, some might argue againstform of prediction, opting instead for a matcing process initiated by the final word. On tview, as a sentence is processed, the seactive features includes those of the curword and of the preceding context words,none of the features of upcoming words.each word is presented, the comprehensiontem presumably checks the degree of (mmatch between context features and the cuword’s features, and responds accordingly.greater the mismatch, the larger the N400 eited. This account would seem to predict tthe more the information in a context constrathe possible exemplars, the greater thematch (and the associated N400) when thatferred exemplar does not occur. While thisreasonable hypothesis, it fails to account forfinding smaller N400s to within-category viotions in high than in low contextual constrai

In contrast, this outcome easily falls outour “prediction” account. The N400 to withicategory violations is reduced because man
gits features, namely those it shares with the

aen

omforasntila-that is400lverts,in

re-maprea

hellycoexth

ione on &rk

ex-ex

iewemsaoningmene?haurare

xperml o

heasatesu

iza

msicalorelu-ureausere-eatetask

inghesed toar-ces

rvedpro-or-

asedicalt totici-allye ofven

ext.ory-un-venoreuc-elyit is

om-

te-te-rt-

text.im-arthatentcepto asin,rstim-


expected (but not presented) exemplar, areready active prior to its appearance. Betwecategory violations share fewer features in cmon with expected exemplars and therecannot benefit from this type of prediction,reflected in a larger N400. It is thus the semarelationship between the within-category viotion and the expected exemplar—the itemwould be predicted in the context but thanever actually seen—that results in an Nreduction (e.g., although pine trees themseare not very compatible with tropical resothey share features in common with someththat is, namely palm trees). Strikingly, thisduction increases when the contextual infortion allows a strong as opposed to a weakdiction of that expected item (andcorrespondingly weaker prediction of twithin-category violation, the item actuaseen). Thus, as a result of processing thetext, the semantic features of the expectedemplar must become activated, and it isoverlap between the within category violatand that prediction that determines the sizthe observed N400 response (see McKooRatcliff, 1989, for a similar conclusion in woon elaborative inferences).

Second, the functional link between unpected but categorically related endings andpected exemplars not only supports the vheld by many researchers that long-term mory is structured but also our specific propothat this structure has an inherent effectsentence processing in real time. Our findthat information is routinely retrieved frolong-term memory during language comprehsion is not news. How could it be otherwisLikewise, we are not the first to suggest tsemantic memory has a categorical structcomponent. The results of categorizationsearch have long been used to infer that erience with the world structures long-tememory, so that items that share perceptuafunctional traits come to be grouped togetand treated as similar to one another, i.e.,category. However, such evidence for the cgorical structure of semantic memory has ually been obtained in some kind of categor
tion task. Participants in these studies are eithe
l---e

c

t

s

g

--

n--

e

f

-

-l

-

tl

--

rra---

explicitly asked to categorize or shown itegrouped in a way that renders their categorrelationship quite apparent. It has therefproven difficult to draw unequivocal concsions from such studies. How can we be sthat the category-based effects arise becindividuals tap into existing structured repsentations and not because individuals crstructure as needed by the categorizationitself (for review, see Kounios, 1996)?

Our results are thus important for resolvthis issue because they are not subject to tconcerns. Our participants were not askeperform any explicit categorization or compison; their only task was to read the sentenfor comprehension. Nevertheless, we obsea reliable category-based effect during thecessing of the sentence final word. It is imptant to note that we observed a category-beffect on the N400, even though the categorrelationship was neither obvious nor relevanthe comprehension task at hand. Our parpants did not even see the two categoricrelated items in close proximity, because onthem—the expected exemplar—was not epresented, but merely implied by the contThus, our results show that robust categbased effects can be obtained within a few hdred milliseconds of an event’s occurrence eoutside of a categorization task. Another, mnovel, finding is that this category-based strture of long term memory seems to routininfluence language processing, even whenirrelevant and perhaps detrimental to the cprehension process.

We observed that the N400 to within-cagory violations is reduced by virtue of its cagorical relationship to the word that is repoedly most expected in the two-sentence conPreviously we argued that neither lexical pring nor plausibility could explain this particulresult, nor could any view that assumeslong-term memory exists without any inherstructure or that there are no categories exthose that are dynamically generated de nova function of context (e.g., Barsalou & Med1986). Our data thus provide one of the ficlear demonstrations that the experientially
rposed structure of long-term memory has a sig-

all

hisicagonocehlyugno

henablonuiksingfa-xtsenn-at-xtsseas

si-tsevklyRPo

amir-

allyconentuof

d bn iforutoro

hextrm

sim-ro-hatbinsn.pleded,

can-hat

ofat&

92;g.,

etdo

tualticend-im-heis

overit

, &5).tivesen-enpsarein

ch,).n offea-gslarup-ey

ges

ludesumeen ats it.anyet ofny ofould


nificant and measurable impact on contextudriven language processes.

Yet another novel finding emerging from tstudy is that the influence of the categorstructure of memory on sentence processinmodulated by the degree to which context cstrains the expected exemplar. Contrary tplausibility or matching account, the influenof this structure actually increases in higconstraining contexts. This observation sgests that the semantic memory structure issimply a factor that becomes relevant wother cues are absent, weak, or less availbut rather that its influence is an inherent csequence of the way the brain processes lingtic input9. Studies using lexical decision tashave previously shown that weakly constraincontexts typically provide a wider scope ofcilitation than do more constraining conte(Schwanenflugel & LaCount, 1988; Schwanflugel & Shoben, 1985). While highly costraining contexts are quite specific in faciliting only the best completion, weaker contecan apparently facilitate a congruent endingmantically related to the best completionwell. We found a similar pattern in the plaubility ratings we obtained off-line: participanwere indeed much more willing to accept seral different endings as plausible in weathan in strongly constraining contexts. The Edata, however, indicate a different patterncontext effects earlier in the processing strearound 300 ms, the apparent facilitation ofrelevant, semantically related items is actugreater in stronger as opposed to weakertexts. On the one hand, high constraint stences seem to provide more specific contexinformation. On the other hand, the effectmemory structure on processing, as indexethe N400, is greater in these contexts thaweaker ones. We take this to mean that inmation about semantic feature overlap is amatically used in language processing, in pportion to the system’s ability to predict tsemantic features of items that will come ne

What does it mean to say that long-te

9 By this we do not mean to imply that it is langua
pecific.
y

lis-a

-t

e,-s-

-

-

-

f;

--al

yn---

.

memory structure (as captured by semanticilarity) has an inherent effect on language pcessing? By this we do not mean to imply tthe brain contains discrete categories orcorresponding to “fruits” or “trees,” and so oAs detailed in the introduction, there is amevidence suggesting that categories are graand overlapping, that category membershipnot be strictly similarity-based, and that wconstitutes a category will vary as a functionboth context and the level of abstractionwhich categorization is performed (BarsalouMedin, 1986; see also review by Komatsu 19Lassaline, Wisniewski, & Medin, 1992; e.Rosch, 1975; Rosch & Mervis, 1975; Roschal., 1976; Roth & Shoben, 1983). What wemean, however, is that the kind of percepand functional similarity captured by semancategories like those we used as sentenceings, but also possibly other types, has anpact on neural and linguistic processing. Tprecise role similarity plays in categorizationunknown and there remains much debatehow best to define similarity or to calculate(e.g., Goldstone, 1995; Medin, GoldstoneGentner, 1993; Murphy & Spalding, 199However, we do know that the brain is sensito various input features and that the repretations of this feature information are oftstructured, for example, into cortical mawherein cells responding to similar featuresphysically close to one another, clusteredcolumnar regions (e.g., Brugge & Merzeni1973; Hubel & Wiesel, 1972; Tanaka, 1996

We can thus view the neural representatiothe objects that words refer to as a set oftures10. Whenever two words refer to two thinthat look alike or sound alike or invoke simimotor programs for interaction, then we spose that there is also a similarity in how th

10 These features need not be simple; they can inchigher order relations. There is also no necessity to asa context-independent or one-to-one mapping betweword and the set of features or neurons that represenThe only critical assumption we make here is that atgiven moment some concept is represented by a sfeatures and that a closely related concept sharing mathose features and activated under similar conditions w
involve activity in a partially overlapping set of neurons.

uraedi-re-g tlly,feaieratuuldurePurethse

tex

lane ton-andme-seom

ordoe. Inrsionfludenn-

renan-dofse

d oa

e on

hatasic-aretex

ithro-we

thee sucht andw omo en-c eser encep andt

BP

A

H

NF

P

V

T

ts


are represented in the brain, i.e., in the neactivity they elicit. A neural system structurin this way would likely find it easier to transtion between the pattern of activation corsponding to one thing and that correspondina different but related thing. More specificaif the comprehension system predicts thetures of “palms” in the way we proposed earlthen a structured representation based on feoverlap of the sort we just described woleave it better prepared to activate the featof “pines” than of “tulips.” Moreover, our ERdata suggest that the more strongly the featof “palms” are activated, the better preparedsystem is to deal with the “pines,” even if thefeatures are a poor fit to the activated confeatures.

In conclusion, our data suggest that theguage comprehension system is sensitivspecific contextual information and to the csistency between that specific informationthe meaning of a target word by around 375into word processing. This information is spcific enough to distinguish two words whoreferents share many semantic features in cmon. However, the fit between a given wand specific contextual information alone dnot completely predict the brain’s responseparticular, in the same time window as we fiobserve the influence of contextual informaton word processing, we also observe an inence of semantic feature overlap (as reflectetaxonomic semantic categories) that is indepdent of the fit of that word to the specific setence context. That is, we observe an inheinfluence of long-term memory structure on lguage processing, at least that aspect indexethe N400. This suggests that the processingsentence context results in the activation of aof semantic features associated with the worwords that are likely to come next. Whenword is actually encountered, it is the degresemantic feature match or mismatch betweeand the prediction derived from context tdetermines the difficulty of processing, at leinitially. Stronger contexts allow better predtions and greater facilitation for items that shfeatures with the predicted word. Thus, con
and long-term memory structure have a dy
l

o

-,re

s

se

t

-o

s

-

s

t

-in-

t

byatr

fit

t

t

namic, mutually dependent relationship wone another and contribute jointly to the pcesses involved in making sense of whatread.

APPENDIX A

Categories Used in the Experiment

Sixty-six different categories were used inxperiment. These categories were paired

hat for two-sentence pairs the expectedithin-category target items were derived frne category of the pair while the betweategory item was derived from the other; tholes reversed for a second set of two-sentairs. The categories used in the experiment

heir pairings were as follows.

iological categorieslantstrees, flowers

nimalscrustaceans, fishmarine mammals, marsupialsdogs, equinesinsects, rodentsbirds, reptiles(wild)cats, bearsdinosaurs, mythical beingsuman-related, human-like itemsbody parts (external), internal organssuperheroes, cartoon animalsonbiological categoriesoodsfruits, vegetablesmeats, cheesesdesserts, breadsalcoholic, non-alcoholic beverages

laces and buildingsland formations, celestial bodiescountries, statesnative dwellings, religious buildings

ehiclescars, public transportationaircraft, boatswar vehicles, heavy machinery

ools and household objectscarpentry tools, gardening toolsmeasuring instruments, optical instrumen

- medical supplies, office supplies

nt

xtswitpela-aretimithviceint

ith

ess

acing

the

n

n-

se

ice

he

rking.es

nd

the

ng

enon-

en

ing

us

the

he

at

st,

he

ed

me.

nd

ave

look

d’s

im,

hat


dishes, utensilssmall (kitchen) appliances, lighting

Garments and personal articlestops, toiletriespants, shoessafety-wear, walking aidesjewelry, make-up

Leisure- and hobby-related itemssports, board gamessports equipment, toysreading material, writing instruments

Miscellaneouscontainers, fasteners

APPENDIX B

Examples of Stimuli Used in the Experime

One hundred thirty-two sentence contewere used in the experiment, each endingone of each of the three possible ending ty(expected exemplars, within-category viotions, between-category violations). Belowgiven 40 representative examples of these suli. Ending types are expected exemplar, win-category violation, and between-categoryolation, respectively. High constraint sentenare marked with an “H” and low constrasentences are marked with an “L.”

(H) “Checkmate,” Rosaline announced wglee.

She was getting to be really good at chmonopoly/football.

(H) Justin put a second house on Park PlHe and his sister often spent hours play

monopoly/chess/baseball.

(H) He caught the pass and scored anotouchdown.

There was nothing he enjoyed more thagood game of football/baseball/monopoly.

(H) Rich couldn’t count the number of Yakees games he had seen with his father.

They both shared a lifelong interest in baball/football/chess.

(L) She felt that she couldn’t leave Venwithout the experience.

It might be a touristy thing to do, but s
wanted to ride in a gondola/ferry/helicopter.
hs

---s

/

e.

r

a

-

(L) Getting both himself and his car to woon the neighboring island was time-consum

Every morning he drove for a few minutand then boarded the ferry/gondola/plane.

(L) The patient was in critical condition athe ambulance wouldn’t be fast enough.

They decided they would have to usehelicopter/plane/ferry.

(L) Amy was very anxious about traveliabroad for the first time.

She felt surprisingly better, however, whshe actually boarded the plane/helicopter/gdola.

(L) The day before the wedding, the kitchwas just covered with frosting.

Annette’s sister was responsible for makthe cake/cookies/toast.

(H) The little girl was happy that Santa Claleft nothing but crumbs on the plate.

She decided he must have really enjoyedcookies/cake/bagel.

(H) Chris moped around all morning whendiscovered there was no cream cheese.

He complained that without it he couldn’t ehis bagel/toast/cake.

(H) He wanted to make his wife breakfabut he burned piece after piece.

I couldn’t believe he was ruining even ttoast/bagel/cookies.

(H) I guess his girlfriend really encouraghim to get it pierced.

But his father sure blew up when he cahome wearing that earring/necklace/lipstick

(L) She keeps twirling it around and arouunder her collar.

Stephanie seems really happy that Dan gher that necklace/earring/mascara.

(H) She wanted to make her eyelashesreally black and thick.

So she asked to borrow her older frienmascara/lipstick/necklace.

(H) He complained that after she kissed hhe couldn’t get the red color off his face.

He finally just asked her to stop wearing t
lipstick/mascara/earring.

e

ea

al

the

y

on

e a

lity

p-stsere

awt a

an

il-

ap-

gedom

red

er

r o

et

eap

he

r o

me

of

ad

her

ail

e to

so

im

a

ke/

theht.

still

or,

big

ned

the

nd

find

cot

e a

a

n


(L) Eleanor offered to fix her visitor somcoffee.

Then she realized she didn’t have a clcup/bowl/spoon.

(L) My aunt fixed my brother some cereusing her best china.

Of course, the first thing he did was dropbowl/cup/knife.

(H) At the dinner party, I wondered why mmother wasn’t eating her soup.

Then I noticed that she didn’t have a spoknife/bowl.

(L) In the dorms, cutting your steak can bhuge struggle.

They always give you such a poor quaknife/spoon/cup.

(H) He journeyed to the African plains, hoing to get a photograph of the king of the bea

Unfortunately, the whole time he was thhe never saw a lion/tiger/panda.

(L) George was hiking in India when he sthe orange and black striped animal leap ouhim.

He sustained serious injuries before he maged to kill the tiger/lion/polar bear.

(H) Hitting the huge animal with a tranquizer dart was difficult in the Arctic winds.

Eventually, however, they were able toproach and tag the polar bear/panda/lion.

(L) Wendy wondered how they had manato ship such a large animal all the way frChina.

She waited in line to see the newly acquipanda/polar bear/tiger.

(H) Barb loved the feel of the waves on hfeet, but she hated to walk barefoot.

As a compromise, she usually wore a paisandals/boots/shorts.

(L) By the end of the day, the hiker’s fewere extremely cold and wet.

It was the last time he would ever buy a chpair of boots/sandals/jeans.

(H) Everyone agreed that the stone-waskind were out of style.

But he continued to wear the same old pai
jeans/shorts/sandals.
n

/

.

t

-

f

d

f

(L) As the afternoon progressed, it becahotter and hotter.

Keith finally decided to put on a pairshorts/jeans/boots.

(L) Pablo wanted to cut the lumber he hbought to make some shelves.

He asked his neighbor if he could borrowsaw/hammer/rake.

(H) Tina lined up where she thought the nshould go.

When she was satisfied, she asked Bruchand her the hammer/saw/shovel.

(H) The snow had piled up on the drivehigh that they couldn’t get the car out.

When Albert woke up, his father handed ha shovel/rake/saw.

(H) The yard was completely covered withthick layer of dead leaves.

Erica decided it was time to get out the rashovel/hammer.

(L) Fred went to the pantry and got outhomemade jelly his grandmother had broug

Fifteen minutes later, however, he wasstruggling to open the jar/box/zipper.

(L) After they unpacked the new refrigeratthey let Billy have his fun.

He played for days afterwards with thebox/jar/button.

(H) It seemed to catch every time she opeor closed her backpack.

She decided she would have to replacezipper/button/box.

(H) One fell off her blouse and got lost, ashe didn’t have any extras.

She ended up searching all over town toa matching button/zipper/jar.

(L) The firefighters wanted to have a masto live with them at the firehouse.

Naturally, they decided it would have to bdalmatian/poodle/zebra.

(L) Muffie, old Mrs. Smith’s pet, wearsbow on the puff of fur on its head.

I don’t know how anyone could want to owa poodle/dalmatian/donkey.

(L) “I’m an animal like Eeyore!” the child
exclaimed.

ing

eyani

ea

ticSp-

B ntiais-

B s oy to-

C 5).andd

D is-earn

D ).ncey:

E oning

F n-ts od

F &en-m-

F ., &ges

F rry,rifi-

F intons

F of

F tialfor

G sim-on.

G ).gerer-l-

G , &pal

H 4).atedantics-

H ofringl

H m-a-y,

K

K 3).uter

n-

K xicalness

K ruc-

K theandtruc-

K cessted

-

K tic.),

st


His mother wondered why he was pretendto be a donkey/zebra/dalmatian.

(H) At the zoo, my sister asked if thpainted the black and white stripes on themal.

I explained to her that they were natural ftures of a zebra/donkey/poodle.

REFERENCES

Barsalou, L. W., & Medin, D. L. (1986). Concepts: Stadefinitions or context-dependent representations?cial Issue: Context and cognition.Cahiers de Psychologie,6, 187–202.

esson, M., & Macar, F. (1987). An event-related poteanalysis of incongruity in music and other non-lingutic contexts.Psychophysiology,24, 14–25.

rugge, J. F., & Merzenich, M. M. (1973). Responseneurons in auditory cortex of the macaque monkemonaural and binaural stimulation.Journal of Neurophysiology,36, 1138–1158.

hao, L. L., Nielsen-Bohlman, L., & Knight, R. T. (199Auditory event-related potentials dissociate earlylate memory processes.Electroencephalography anClinical Neurophysiology,96, 157–168.

ale, A. M. (1994).Source localization and spatial dcriminant analysis of event-related potentials: linapproaches.La Jolla, CA: University of California SaDiego.

uffy, S. A., Henderson, J. M., & Morris, R. K. (1989Semantic facilitation of lexical access during senteprocessing. Journal of Experimental PsychologLearning, Memory, & Cognition,15, 791–801.

hrlich, S. F., & Rayner, K. (1981). Contextual effectsword perception and eye movements during readJournal of Verbal Learning and Verbal Behavior,20,641–655.

ischler, I., & Bloom, P. A. (1979). Automatic and attetional processes in the effects of sentence contexword recognition.Journal of Verbal Learning anVerbal Behavior,18, 1–20.

ischler, I., Bloom, P. A., Childers, D. G., Arroyo, A. A.,Perry, N. W. J. (1984). Brain potentials during stence verification: Late negativity and long-term meory strength.Neuropsychologia,22, 559–568.

ischler, I., Bloom, P. A., Childers, D. G., Roucos, S. EPerry, N. W. (1983). Brain potentials related to staof sentence verification.Psychophysiology,20, 400–409.

ischler, I., Childers, D. G., Achariyapaopan, T., & PeN. W. (1985). Brain potentials during sentence vecation: Automatic aspects of comprehension.Biologi-cal Psychology,21, 83–105.

ischler, I. S., & Bloom, P. A. (1985). Effects of constraand validity of sentence contexts on lexical decisi
Memory & Cognition,13, 128–139. K
-

-

e-

l

f

.

n

.

rancis, W. N., & Kucera, H. (1982).Frequency analysisEnglish usage.Boston: Houghton Mifflin.

riedman, D. (1990). Cognitive event-related potencomponents during continuous recognition memorypictures.Psychophysiology,27, 136–148.

oldstone, R. L. (1995). Mainstream and avant-gardeilarity. Special Issue: Similarity and categorizatiPsychologica Belgica,35, 145–165.

ough, P. B., Alford, J. A., Jr., & Holley-Wilcox, P. (1981Words and contexts. In O. J. L. Tzeng & H. Sin(Eds.),Perception of print: Reading research in expimental psychology(pp. 85–102). Hillsdale, NJ: Erbaum.

runwald, T., Elger, C. E., Lehnertz, K., Van Roost, D.Heinze, H. J. (1995). Alterations of intrahippocamcognitive potentials in temporal lobe epilepsy.Electro-encephalography and Clinical Neurophysiology,95,53–62.

arbin, T. J., Marsh, G. R., & Harvey, M. T. (198Differences in the late components of the event-relpotential due to age and to semantic and non-semtasks.Electroencephalography & Clinical Neurophyiology: Evoked Potentials,59, 489–496.

ess, D. J., Foss, D. J., & Carroll, P. (1995). Effectsglobal and local context on lexical processing dulanguage comprehension.Journal of ExperimentaPsychology: General,124,62–82.

ubel, D. H., & Wiesel, T. N. (1972). Laminar and colunar distribution of geniculo-cortical fibers in the mcaque monkey.Journal of Comparative Neurolog146,421–450.

ay, P. (1971). Taxonomy and semantic contrast.Lan-guage,47, 866–887.

iss, G. R., Armstrong, C., Milroy, R., & Piper, J. (197An associated thesaurus of English and it companalysis. In A. J. Aitken, R. Bailey, & N. HamiltoSmith (Eds.),The computer and literary studies.Ed-inburgh: Edinburgh University Press.

leiman, G. M. (1980). Sentence frame contexts and ledecisions: Sentence-acceptability and word-relatedeffects.Memory & Cognition,8, 336–344.

omatsu, L. K. (1992). Recent views of conceptual stture.Psychological Bulletin,112,500–526.

ounios, J. (1996). On the continuity of thought andrepresentation of knowledge: Electrophysiologicalbehavioral time-course measures reveal levels of sture in semantic memory.Psychonomic Bulletin &Review,3, 265–286.

ounios, J., & Holcomb, P. J. (1992). Structure and proin semantic memory: Evidence from event-relabrain potentials and reaction times.Journal of Experimental Psychology: General,121,459–479.

utas, M., & Dale, A. (1997). Electrical and magnereadings of mental functions. In M. D. Rugg (EdCognitive Neuroscience(pp. 197–242). Hove, EaSussex: Psychology Press.

utas, M., & Hillyard, S. A. (1980a). Event-related brain

ris.essgru

inma

ngsoc

).tialRe

ticsns-.2).all

Perd

).

l of&

g:J:

D.an

ion

ofwity

nssua-or-

onsl

3).

M nc

og-

M e,

L.tialande,

ted: II.

is of-

tego--

-

R of

R anticn-

R es:

R , &cat-

R tual

R texty,

Rnd

s.S 1).

the

S nticingl-

S encefor

e,

S d in-re of-


potentials to semantically inappropriate and surpingly large words.Biological Psychology,11,99–116

Kutas, M., & Hillyard, S. A. (1980b). Reading senselsentences: Brain potentials reflect semantic inconity. Science,207,203–205.

Kutas, M., & Hillyard, S. A. (1983). Event-related brapotentials to grammatical errors and semantic anolies. Memory & Cognition,11, 539–550.

Kutas, M., & Hillyard, S. A. (1984). Brain potentials durireading reflect word expectancy and semantic asation.Nature,307,161–163.

Kutas, M., Lindamood, T. E., & Hillyard, S. A. (1984Word expectancy and event-related brain potenduring sentence processing. In S. Kornblum & J.quin (Eds.),Preparatory States and Processes(pp.217–237). Hilldale, NJ: Erlbaum.

Kutas, M., & Van Petten, C. K. (1994). Psycholinguiselectrified: Event-related brain potential investigatioIn M. A. Gernsbacher (Ed.),Handbook of psycholinguistics(pp. 83–143). San Diego: Academic Press

Lassaline, M. E., Wisniewski, E. J., & Medin, D. L. (199Basic levels in artificial and natural categories: Arebasic levels created equal? In B. Barbara (Ed.),cepts, concepts and categories:The representation anprocessing of information.(Vol. 93, pp. 327–378Amsterdam: North-Holland.

Masson, M. E. J. (1991). A distributed memory modecontext effects in word identification. In D. BesnerG. W. Humphreys (Eds.),Basic processes in readinVisual word recognition(pp. 233–263). Hillsdale, NErlbaum.

McCarthy, G., Nobre, A. C., Bentin, S., & Spencer, D.(1995). Language-related field potentials in theterior-medial temporal lobe: I. Intracranial distributand neural generators.Journal of Neuroscience,15,1080–1089.

McCarthy, G., & Wood, C. C. (1985). Scalp distributionsevent-related potentials: an ambiguity associatedanalysis of variance models.Electroencephalographand Clinical Neurophysiology,62, 203–208.

McClelland, J. L., & O’Regan, J. K. (1981). Expectatioincrease the benefit derived from parafoveal viinformation in reading words aloud.Journal of Experimental Psychology: Human Perception & Perfmance,7, 634–644.

McKoon, G., & Ratcliff, R. (1989). Semantic associatiand elaborative inference.Journal of ExperimentaPsychology: Learning, Memory, & Cognition,15,326–338.

Medin, D. L., Goldstone, R. L., & Gentner, D. (199Respects for similarity.Psychological Review,100,254–278.

orris, R. K. (1994). Lexical and message-level sentecontext effects on fixation times in reading.Journal ofExperimental Psychology: Learning, Memory, & Cnition, 20, 92–103.

urphy, G. L., & Spalding, T. L. (1995). Knowledg
similarity, and concept formation. Special Issue: Sim-
-

-

-

i-

s-

.

-

-

h

l

e

ilarity and categorization.Psychologica Belgica,35,127–144.

Neville, H. J., Kutas, M., Chesney, G., & Schmidt, A.(1986). Event-related brain potentials during iniencoding and recognition memory of congruousincongruous words.Journal of Memory and Languag25, 75–92.

Nobre, A. C., & McCarthy, G. (1995). Language-relafield potentials in the anterior-medial temporal lobeEffects of word type and semantic priming.Journal ofNeuroscience,15, 1090–1098.

Oldfield, R. C. (1971). The assessment and analyshandedness: The Edinburgh inventory.Neuropsychologia, 9, 97–113.

Polich, J. (1985). N400s from sentences, semantic caries, number and letter strings?Bulletin of the Psychonomic Society,23, 361–364.

Ratcliff, J. E. (1987). The plausibility effect: Lexical priming or sentential processing?Memory & Cognition,15,482–496.

atcliff, R., & McKoon, G. (1988). A retrieval theorypriming in memory.Psychological Review,95, 385–408.

osch, E. (1975). Cognitive representations of semcategories.Journal of Experimental Psychology: Geeral, 104,192–233.

osch, E., & Mervis, C. B. (1975). Family resemblancStudies in the internal structure of categories.Cogni-tive Psychology,7, 573–605.

osch, E., Mervis, C. B., Gray, W. D., Johnson, D. M.Boyes-Braem, P. (1976). Basic objects in naturalegories.Cognitive Psychology,8, 382–439.

osch, E. H. (1973). On the internal structure of percepand semantic categories. In T. E. Moore (Ed.),Cogni-tive development and the acquisition of language(Vol.xii, p. 308). New York: Academic Press.

oth, E. M., & Shoben, E. J. (1983). The effect of conon the structure of categories.Cognitive Psycholog15, 346–378.

ugg, M. D., & Coles, M. G. H. (Eds.). (1995).Electro-physiology of mind: Event-related brain potentials acognition. (Vol. 25). Oxford: Oxford University Pres

chuberth, R. E., Spoehr, K. T., & Lane, D. M. (198Effects of stimulus and contextual information onlexical decision process.Memory & Cognition,9, 68–77.

chwanenflugel, P. J., & LaCount, K. L. (1988). Semarelatedness and the scope of facilitation for upcomwords in sentences.Journal of Experimental Psychoogy: Learning, Memory, & Cognition,14, 344–354.

chwanenflugel, P. J., & Shoben, E. J. (1985). The influof sentence constraint on the scope of facilitationupcoming words.Journal of Memory and Languag24, 232–252.

chwantes, F. M. (1985). Expectancy, integration, anteractional processes: Age differences in the natuwords affected by sentence context.Journal of Exper
imental Child Psychology,39, 212–229.

6).in

y

S

S y ay:

S ch-ten

T ion

T for

V (indis-

V eenlated

Z ring

(


Smith, M. E., Stapleton, J. M., & Halgren, E. (198Human medial temporal lobe potentials evokedmemory and language tasks.Electroencephalographand Clinical Neurophysiology,63, 145–159.

quire, L. R. (1987).Memory and brain.New York: OxfordUniversity Press.

tanovich, K. E., & West, R. F. (1983). On priming bsentence context.Journal of Experimental PsychologGeneral,112,1–36.

tuss, D. T., Picton, T. W., & Cerri, A. M. (1986). Searing for the names of pictures: An event-related potial study.Psychophysiology,23, 215–223.

anaka, K. (1996). Inferotemporal cortex and object vis
Annual Review of Neuroscience,19, 109–139. (
-

.

aylor, W. L. (1953). “Cloze procedure”: a new toolmeasuring readability.Journalism Quarterly,30,415–433.

an Berkum, J. J. A., Hagoort, P., & Brown, C. M.press). Semantic integration in sentences andcourse.Journal of Cognitive Neuroscience.

an Petten, C., & Kutas, M. (1990). Interactions betwsentence context and word frequency in event-rebrain potentials.Memory & Cognition,18, 380–393.

ola, D. (1984). Redundancy and word perception dureading.Perception & Psychophysics,36, 277–284.

Received March 5, 1998)
Revision received April 13, 1999)

Documents

A Rose by Any Other Name: Long-Term Memory Structure and ...ling.umd.edu/~ellenlau/courses/ling848/Federmeier_1999.pdf · A Rose by Any Other Name: Long-Term Memory Structure and