Paintings & Photography

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

A Rose by Any Other Name: Long-Term Memory Structure and Sentence Processing
of 27
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
  A Rose by Any Other Name: Long-Term Memory Structureand Sentence Processing Kara D. Federmeier  Department of Cognitive Science, University of California, San Diego andMarta 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. At its heart, language comprehension involvesthe recruitment and integration of world knowl-edge stored in long-term memory. Consider, forexample, 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 theyexpect the missing final word of this sentence pairto be “ferry.” How do they come to that expecta-tion? None of the individual words in this sen-tence pair is strongly associated with the word  ferry. There are, in fact, a number of differentvehicle names that could plausibly complete thesentence. Yet, the expectation is remarkably con-sistent across individuals. In order to create theirexpectations, readers must use information in thesentence context to build a cognitive model in-volving vehicles that can transport both peopleand cars across water and that are likely to be usedhabitually. It is their store of world knowledge,combined with this model, that allows readers tothen determine that the vehicle in question islikely to be a ferry and not an ocean liner, barge,airplane, or helicopter. Given how crucial long-term memory is for language processing, it issomewhat surprising that so little is known abouthow information from memory is accessed andused during on-line language processing. Context Effects in Language Off-line tasks make it clear that there is oftensufficient information in a sentence context to This work was supported by a Howard Hughes Predoc-toral Fellowship to K.F. and grants HD22614, AG08313,and MH52893 to M.K. We thank Elizabeth Bates, SeanaCoulson, Phillip Holcomb, Jonathan King, David Kirsh,Marty Sereno, and two anonymous reviewers for helpfulcomments and discussion.Address reprint requests to Kara D. Federmeier, Depart-ment of Cognitive Science, University of California, SanDiego, 9500 Gilman Drive, La Jolla, CA 92093-0515.E-mail:; Fax: 619-534-1128.4690749-596X/99 $30.00 Copyright © 1999 by Academic PressAll rights of reproduction in any form reserved. Journal of Memory and Language 41, 469–495 (1999)Article ID jmla.1999.2660, available online at on  significantly constrain guesses about what con-cepts or even words are likely to be next en-countered. However, whether—and, if so, whenand how—this information affects processingon-line remains a hotly debated issue. Muchpsycholinguistic research suggests that wordsthat are predictable in a sentence context areperceived and processed more rapidly and ac-curately than the same words when they occurout of context or in incongruent contexts. Forexample, contextual information decreases theduration of readers’ eye fixations (Ehrlich &Rayner, 1981; Morris, 1994; Zola, 1984). Con-gruent contexts also facilitate the time to pro-nounce sentence-final or phrase-final words(Duffy, Henderson, & Morris, 1989; Hess, Foss,& Carroll, 1995; McClelland & O’Regan, 1981;Stanovich & West, 1983) and the speed of word/nonword judgments (lexical decision) onthem (Fischler & Bloom, 1985; Kleiman, 1980;Schuberth, Spoehr, & Lane, 1981). This facili-tation occurs even when the relatedness of lex-ical items within congruent and incongruentsentences is matched, suggesting that the ob-served increase in processing fluency cannot beattributed solely to lexical priming, but involvesinformation provided by the sentence as a whole(e.g., Duffy et al., 1989; Morris, 1994; Ratcliff,1987).Electrophysiological results support thesefindings and suggest that contextual informationis used early and builds continuously over thecourse of processing a sentence. The event-related brain potential (ERP) technique involvesrecording at the scalp neural activity that istime-locked to a particular event. The neuralactivity recorded is known to reflect the sum-mation of graded postsynaptic potentials, pre-dominantly from pyramidal cells of the cerebralcortex (for review, see Kutas & Dale, 1997).The ERP technique provides a continuous, mul-tidimensional measure that can be recorded dur-ing natural language processing (without theimposition of an additional task). It providesmillisecond-level temporal resolution and infor-mation about the number and, in some cases,location of the neural sources contributing to agiven task or condition (e.g., Rugg & Coles,1995).An ERP component that has proven espe-cially useful for the study of contextual influ-ences in language processing is the N400, anegative-going potential peaking around 400ms after stimulus onset. Kutas and Hillyard(1980b) first observed the N400 during a task inwhich individuals read sentences word by wordfor comprehension. Sentence final words thatwere semantically anomalous with respect tothe sentence context were associated with asignificantly larger negativity 250 to 600 mspost-stimulus-onset than were words that fit thesentence context. Subsequent investigationshave revealed that each word in a sentenceelicits an N400 and that the amplitude of thiscomponent is highly correlated with individu-als’ off-line expectations as measured by “clozeprobability 1 ” (Kutas & Hillyard, 1984) and de-creases as contextual information builds overthe course of a sentence (Van Petten & Kutas,1990).The influence of contextual information onword processing has been most clearly demon-strated for words that are highly predictable intheir sentence contexts (“best completions”; i.e.,words with the highest cloze probability in thecontext). However, to a more limited degree,contextual information has also been found toaffect the processing of less predictable words.With behavioral techniques, for example, someresearchers find facilitation for unexpected butcontextually congruous words (e.g., Schwantes,1985; Stanovich & West, 1983); others, how-ever, do not (e.g., Fischler & Bloom, 1979;Kleiman, 1980; Schwanenflugel & LaCount,1988). In electrophysiological investigations,these congruent but low cloze probability itemselicit N400 responses that are larger than thoseto higher cloze probability items but smallerthan those to contextually incongruent items(e.g., Kutas & Hillyard, 1984). Both behavioraland electrophysiological studies have also ob-served facilitation for unexpected items that aresemantically related to the best completion(Kleiman, 1980; Kutas & Hillyard, 1984; Kutas, 1 The cloze probability of a word in a given context refersto the proportion of people who would choose to completethat particular sentence fragment with that particular word(Taylor, 1953). 470 FEDERMEIER AND KUTAS  Lindamood, & Hillyard, 1984; Schwanenflugel& LaCount, 1988); these effects can be ob-served even for words that do not form accept-able sentence completions (Kleiman, 1980; Ku-tas & Hillyard, 1984; Kutas et al., 1984).The types of words facilitated by a contextand the degree of facilitation for each seem tovary with the nature of the context itself. Forexample, highly constraining contexts seem toprovide greater facilitation of “best comple-tions” than do less constraining contexts(Fischler & Bloom, 1979; McClelland &O’Regan, 1981). At the same time, however,highly constraining contexts have a narrower“scope” of facilitation that does not extend toless predictable items (e.g., Schwanenflugel &LaCount, 1988; Schwanenflugel & Shoben,1985). Less constraining contexts, on the otherhand, facilitate a wider range of items and pro-vide enhanced facilitation for less predictableitems. Research has also shown that withgreater semantic-associative information (i.e.,more words that are lexically associated with atarget) in a context, one observes greater facil-itation of contextually congruent words (Duffyet al., 1989) and more elaborative inference-drawing (McKoon & Ratcliff, 1989); this factorhas not always been controlled for in otherstudies looking at, for example, effects of con-text on contextually incongruent, semanticallyassociated targets.Taken together, this body of work suggeststhat sentence contexts facilitate, in a gradedmanner, the processing of a set of conceptsand/or words. Moreover, the nature and strengthof the sentence context affects what items/con-cepts are facilitated and to what extent. How-ever, it cannot be information in the sentencecontext alone 2 that determines what is or is notfacilitated, as at times facilitation has been ob-served for contextually inappropriate (but se-mantically related) items but not observed forunexpected but contextually congruent ones(Schwanenflugel & LaCount, 1988). Becauselanguage comprehension crucially relies on in-formation stored in long-term memory, we hy-pothesized that the structured nature of thismemory is another significant—but relativelyunexplored—variable likely to be affecting howwords are processed during reading. The N400 and Long-Term Memory The hypothesis that the organization of se-mantic memory plays an integral role in deter-mining how information in a sentence contextwill affect word processing receives supportfrom the observation that N400 effects, whileinsensitive to nonsemantic manipulations of context (e.g., changes in the physical attributesof words (Kutas & Hillyard, 1980a) or gram-matical and morphological violations (Kutas &Hillyard, 1983)) or deviations in nonlinguisticstimuli (e.g., anomalous notes in melodies (Bes-son & Macar, 1987)), do seem to be sensitive tolong-term memory processes. N400 compo-nents have been recorded during investigationsof recognition memory for both words (Neville,Kutas, Chesney, & Schmidt, 1986; Smith,Stapleton, & Halgren, 1986) and pictures(Friedman, 1990). Some studies have claimedto observe N400-like components during mem-ory tasks involving stimuli that are not particu-larly semantic in nature. For example, Stuss etal. (1986) reported an N400-like componentwhose amplitude varied with the number of pictures to be remembered in a continuous rec-ognition–memory task. Chao et al. (1995) alsoreported what they describe as an N400 effect toenvironmental noise stimuli, but only duringconditions involving repetitions after long de-lays. Both groups suggest that their findingsimplicate the N400 component in searchthrough long-term memory.Studies into the neurobiological basis of theN400 effect also support a link between thiscomponent and long-term memory processes.McCarthy et al. (1995; also Nobre & McCarthy,1995) recorded field potentials from intracranialelectrodes implanted in humans undergoingtreatment for epilepsy as they read sentences forcomprehension. Anomalous sentence endingswere associated with large field potentials in the 2 Of course, no sentential context effects are wholly in-dependent of long-term memory, as context effects neces-sarily derive from information stored in memory. When wespeak of the influence of sentence context alone, we refer tothe kinds of sentence context effects that would be expectedeven if memory were unstructured. 471 LONG-TERM MEMORY STRUCTURE AND SENTENCE PROCESSING  left and right anterior medial temporal lobes.The authors suggested that these potentials weregenerated in anterior fusiform and parahip-pocampal gyri and perhaps the hippocampusproper. Grunwald et al. (1995) also showed thatthe presence or absence and amplitude of fieldpotentials recorded from the left anterior medialtemporal lobe were correlated with performanceon a delayed word recognition task. As medialtemporal lobe structures are considered criticalfor successful performance on declarative mem-ory tasks (e.g., Squire, 1987), a medial temporallobe source for at least some of the N400 ac-tivity at the scalp lends credence to the idea thatthe N400 may index semantic memory involve-ment as a word is integrated with previous con-text.Insofar as context effects—and associatedN400 effects—derive from perceivers’ knowl-edge about the world and the access of thatknowledge from memory, on-line language pro-cessing should be influenced by the structure of that memory. The structure of semantic memorymay be based on many factors, but it seemslikely that an important part of its organizationcould involve the kind of featural similaritystructure that has been observed to underliehuman categorization (and information repre-sentation in the brain; see, e.g., Tanaka, 1996,for a striking example from higher order visualrepresentation). Categorization research sug-gests that many human categories are taxo-nomic : items are grouped together on the basisof shared perceptual and functional attributes(e.g., Kay, 1971; Rosch, Mervis, Gray, Johnson,& Boyes-Braem, 1976) and these groupings oc-cur at multiple levels of generality, similar tobiological taxonomies. In this scheme, categorymembership is graded, determined by whether—and how many—attributes an item shareswith other members of a category (e.g., Rosch,1975; Rosch & Mervis, 1975; Rosch, 1973). Inturn, there also seems to be a structured, gradedorganization of categories themselves. For ex-ample, a particular item may be called “a plant,”“a flower,” or “a rose”—each a category itself,albeit with successively decreasing inclusive-ness.Consistent with a relationship between taxo-nomic categories, long-term memory, and con-text effects, electrophysiological studies haveshown that N400 amplitude is sensitive to cat-egory membership. For example, in studies byPolich (1985) and Harbin et al. (1984) volun-teers were shown a series of words belonging toa particular taxonomic category. The word se-ries ended either with another member of thecategory or with a word from a different cate-gory. Both groups found that a final word thatwas not a category member generated moreN400 activity than did a category member. Cat-egory effects on the N400 have also been ob-served during the performance of sentence ver-ification tasks (in which individuals are asked to judge the truth of statements of the form “An X  is a Y  ”—for example, “A robin is a bird”) (e.g.,Fischler, Bloom, Childers, Roucos, & Perry,1983; Kounios, 1996; Kounios & Holcomb,1992). In these tasks, an enlarged N400 re-sponse is observed to the final word of falsestatements such as “A carrot is a fruit.” Reveal-ingly, in some cases large N400 responses arealso observed at the end of true statements suchas “A carrot is not a fruit” (e.g., Fischler et al.,1983). In other words, at least in some contexts,the categorical relationship between the subjectand object actually seems to be a more reliablepredictor of the N400 response than the fit of theitem in the sentence context itself (though ef-fects of propositional truth on the N400 havealso been reported (Fischler, Bloom, Childers,Arroyo, & Perry, 1984; Fischler, Childers,Achariyapaopan, & Perry, 1985)). The Present Study Taken together, behavioral and electrophysi-ological evidence suggest that the impact of asentence context on a word’s processing may beinfluenced by and interact with the structure of knowledge in long-term memory—a structurethat is likely based, at least in part, on theperceptual and functional similarity captured bysemantic categories. In fact, the first influencesof both semantic context and category member-ship on lexical processing are manifest as achange in the amplitude of the same ERP com-ponent—the N400. Therefore, we can use theN400 to examine the extent to which long-term472 FEDERMEIER AND KUTAS  memory structure interacts with contextual in-formation during on-line sentence processing.In particular, in this study we examined whetherreaders’ processing would be affected by mem-ory structure even when that structure was ir-relevant to the language comprehension task. Inaddition, we aimed at getting a better under-standing of the role of memory structure inreading by comparing its influence when sen-tence contexts are strong versus when they areweaker.We addressed these issues by comparing theeffects of two types of contextual violations: (1)those that come from the same semantic cate-gory as the contextually predicted item and thusshare many features in common with it(“within-category violations”) and (2) those thatcome from different semantic categories andthus share far fewer features in common withthe predicted item (“between-category viola-tions”). ERPs were recorded as volunteers readpairs of sentences. Each sentence pair was de-signed to create an expectation for a specificexemplar of a specific category (e.g., “Theywanted to make the hotel look more like atropical resort. So along the driveway theyplanted rows of . . . ”). The second sentence of the pair ended with either (1) the expected ex-emplar (“palms”), (2) an unexpected exemplarfrom the same category as the expected exem-plar (“pines”), termed the within-category vio-lation, or (3) an unexpected exemplar from adifferent category than the expected exemplar(“tulips”), termed the between-category viola-tion. It is important to note that exemplars forthe between-category violations were still mem-bers of a shared higher-level category (e.g.,“plants”) and therefore match the other twoitems on most general dimensions. Items rotatedroles across sentences such that they served aseach type of ending once across the stimulusset, as the following example illustrates:1. They wanted to make the hotel look morelike a tropical resort. So along the driveway,they planted rows of  palms/pines/tulips. 2. The air smelled like a Christmas wreathand the ground was littered with needles. Theland in this part of the country was just coveredwith pines/palms/roses. 3. The gardener really impressed his wife onValentine’s Day. To surprise her, he had se-cretly grown some roses/tulips/palms. 4. The tourist in Holland stared in awe at therows and rows of color. She wished she lived ina place where they grew tulips/roses/pines. The sentence contexts varied in their constraint,the degree to which they led to a strong (con-sistent) expectation for the best completion.Comparing the pattern of ERP results ob-tained when individuals read sentences with andwithout violations of the two types should helpunravel the importance of sentence context in-formation and semantic memory structure forlanguage comprehension. Previous work sug-gests that the best completions (i.e., the ex-pected exemplars) will elicit a positivity be-tween 300 and 500 ms. By contrast, thebetween-category violations, which are contex-tually unexpected, difficult to integrate, andshare few features in common with the bestcompletions, will likely elicit an N400 in thesame time window (e.g., Kutas & Hillyard,1980b). What is unknown from previous work is how the response to the within-category vio-lations will compare with the response to ex-pected exemplars and between-category viola-tions.Within-category violations are similar to ex-pected exemplars in that they share many se-mantic features in common. Therefore, if (at thelevel of processing indexed by the N400) thesystem is sensitive only to a fairly general fea-ture match between an item and a sentencecontext, one might expect a similar amplitude toexpected exemplars and within-category viola-tions. A difference between expected exemplarsand within-category violations would suggestthat the system is sensitive to more specificcontextual information (the kind that allows in-dividuals, off-line, to predict the expected ex-emplar but not the within-category violation).Alternatively, if the system is sensitive tospecific contextual information—and thatalone—one would expect an N400 of similaramplitude to both within- and between-category473 LONG-TERM MEMORY STRUCTURE AND SENTENCE PROCESSING
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks