2.28 Semantic Memory - web.colby.eduweb.colby.edu/memoryandlanguagelab/files/2012/07/Balota-Coane-2008.pdf2.28.5 Grounding Semantics 520 2.28.5.1 Grounding Semantics in Analyses of

Author's personal copy

2.28 Semantic MemoryD. A. Balota and J. H. Coane, Washington University, St. Louis, MO, USA

ª 2008 Elsevier Ltd. All rights reserved.

2.28.1 Nature of the Representation 5122.28.2 Network Approaches 5122.28.3 Feature Analytic Approaches 5162.28.4 Concept Learning and Categorization 5182.28.5 Grounding Semantics 5202.28.5.1 Grounding Semantics in Analyses of Large-Scale Databases 5202.28.5.2 Grounding Semantics in Perceptual Motor Systems 5212.28.6 Measuring Semantic Representations and Processes: Insights from Semantic

Priming Studies 5222.28.7 The Interplay Between Semantics and Episodic Memory 5252.28.8 Representation and Distinctions: Evidence from Neuropsychology 5272.28.8.1 Category-Specific Deficits 5272.28.8.2 Semantic Dementia 5282.28.9 Neuroimaging 5292.28.10 Development and Bilingualism 5312.28.11 Closing Comments 531References 531

Semantic memory entails the enormous store-house of knowledge that all humans have available.To begin with, simply consider the informationstored about the words of one’s native language.Each of us has approximately 50,000 words storedin our mental dictionary. With each entry, we alsohave many different dimensions available. For exam-ple, with the word ‘dog’ we have stored informationabout how to spell it, how to pronounce it, its gram-matical category, and the fact that the object the wordrefers to typically has four legs, is furry, is a commonpet, and likes to chase cats (sometimes cars, squirrels,and other rodents), along with additional sensoryinformation about how it feels when petted, thesound produced when it barks, the visual appearanceof different types of dogs, emotional responses frompast experiences, and much, much more. Of course,our knowledge about words is only the tip of theiceberg of the knowledge we have available. Forexample, people (both private and public) are a par-ticularly rich source of knowledge. Consider howeasy it is to quickly and efficiently retrieve detailedcharacteristics about John F. Kennedy, MarilynMonroe, Bill Clinton, a sibling, parent, child, and soon. Indeed, our semantic, encyclopedic knowledgeabout the world appears limitless.

One concern reflected by the examples above is thatsemantic memory seems to be all inclusive. In thislight, it is useful to contrast it with other forms ofmemory, and this is precisely what Tulving (1972)did in his classic paper distinguishing semantic andepisodic memory. According to Tulving, semanticmemory ‘‘is a mental thesaurus, organized knowledgea person possesses about words and other verbal sym-bols, their meaning and referents, about relationsamong them, and about rules, formulas, and algorithmsfor the manipulation of these symbols, concepts andrelations’’ (1972, p. 386). In contrast, episodic memoryrefers to a person’s memory for specific events thatwere personally experienced and remembered. So,the memory for the experience of having breakfastyesterday (e.g., where one was seated, how one felt,the taste of the food, who one was with) would fallunder the umbrella of episodic memory, but the factthat eggs, cereal, and toast are typical breakfast foodsreflects semantic knowledge. However, as we shall see,there is some controversy regarding where episodicmemory ends and semantic memory begins. Indeed,we would argue that semantic memory penetrates allforms of memory, even sensory and working memory(Sperling, 1960; Tulving and Pearlstone, 1966;Baddeley, 2000), because tasks that are assumed to

511


tap into these other types of memory often are influ-enced by semantic memory.

So, what is indeed unique about semantic memory,and how has this area of research contributed to ourunderstanding of learning and memory in general?One issue that researchers in this area have seriouslytackled is the nature of representation, which toucheson issues that have long plagued the philosophy ofknowledge or epistemology. Specifically, what does itmean to know something? What does it mean torepresent the meaning of a word, such as DOG? Is itsimply some central tendency of past experienceswith DOGS that one has been exposed to (i.e., aprototype DOG), or is there a limited list of primitivesemantic features that humans use to capture themeaning of DOG, along with many other conceptsand objects? Is the knowledge stored in an abstracted,amodal form that is accessible via different routes orsystems, or is all knowledge grounded in specificmodalities? For example, the meaning of DOGmight be represented by traces laid down by theperceptual motor systems that were engaged whenwe have interacted with DOGs in the past.

In this chapter, we attempt to provide an overviewof the major areas of research addressing the nature ofsemantic memory, emphasizing the major themesthat have historically been at the center of research.Clearly, given the space limitations, the goal here isto introduce the reader to these issues and providereferences to more detailed reviews. The vast major-ity of this work emphasizes behavioral approaches tothe study of semantics, but we also touch upon con-tributions from neuropsychology, neuroimaging, andcomputational linguistics that have been quite infor-mative recently. We focus on the following majorhistorical developments: (1) the nature of the repre-sentation, (2) conceptual development and learning,(3) insights from and limitations of semantic primingstudies, (4) interplay between semantic and episodicmemory tasks, and (5) cognitive neuroscience con-straints afforded by comparisons of different patientpopulations and recent evidence from neuroimagingstudies. For further discussion of this latter area, theinterested reader should see Chapter 2.29.

2.28.1 Nature of the Representation

Although the question of how one represents knowl-edge has been around since the time of Aristotle, itis clear that cognitive scientists are still activelypursuing this issue. One approach to representation

is that we abstract from experience a prototypicalmeaning of a concept, and these ideal representationsare interconnected to other related representationswithin a rich network of semantic knowledge. This isthe network approach. Another approach is that thereis a set of primitive features that we use to define themeaning of words. The meanings of different wordsand concepts reflect different combinations of theseprimitive features. This is a feature-based approach.Historically, the distinction between these twoapproaches has been central to research addressingthe nature of semantic memory.

2.28.2 Network Approaches

One of the first landmark studies of knowledge rep-resentation came from computer science and wasbased on the important dissertation of A. M.Quillian. Quillian (1968) developed a model ofknowledge representation called the TeachableLanguage Comprehender. A goal of this model wasto formulate a working program that allowed effi-cient access to an enormous amount of informationwhile minimizing redundancy of information in thenetwork. Quillian adopted a hierarchically organizednetwork, a portion of which is displayed in Figure 1.As shown, there are two important aspects to thenetwork: nodes and pathways. The nodes in this net-work are intended to directly represent a concept insemantic memory, so for example, the word BIRDhas a node that represents BIRDNESS. These nodesare interconnected in this network via labeled path-ways, which are either ‘isa’ directional pathways orproperty pathways. Specifically, one can verify thatBIRDS are indeed ANIMALS by finding an isapathway between BIRDS and ANIMALS. Likewise,one could verify that ‘A ROBIN BREATHES’ byfinding the isa pathway between robin and bird, andbetween bird and animal, and then accessing theproperty pathway leading to BREATHES fromANIMALS. In this sense, the model was quite eco-nomical, because most properties were stored only atthe highest level in the network in which most of thelower exemplars included that property. For exam-ple, BREATHES would only be stored at theANIMAL level, and not at the BIRD or CANARYlevel, thereby minimizing redundancy (and memorystorage) in the network. Quillian also recognized thatsome features may not apply to all exemplars belowthat level in the network (e.g., ostriches are birds, andbirds fly), so in these cases, one needed to include a

512 Semantic Memory


special property for these concepts (such as CAN’TFLY attached to OSTRICHES).

The economy of the network displayed inFigure 1 does not come without some cost.Specifically, why would one search so deeply in anetwork to verify a property of a given concept, thatis, why would one have to go all the way to theANIMAL concept to verify that ‘CANARIESBREATHE’? It seems more plausible that we wouldhave the property BREATHES directly stored withthe CANARY node. Of course, Quillian was notinitially interested in how well his network mightcapture performance in humans, because his goalwas to develop a computer model that would beable to verify a multitude of questions about naturalcategories, within the constraints of precious compu-ter memory available at the time.

Fortunately for cognitive psychologists, Quillianbegan a collaborative effort with A. Collins to testwhether the network model developed by Quilliancould indeed predict human performance on a sen-tence verification task, that is, the speed to verifysuch sentences as ‘A CANARY IS A BIRD’.Remarkably, the Collins and Quillian (1969) studyprovided evidence that appeared to be highly sup-portive of the hierarchically organized networkstructure that Quillian independently developed inartificial intelligence. Specifically, human perfor-mance was nicely predicted by how many ‘isa’ and

‘property’ pathways one needed to traverse to verifya sentence. The notion is that there was a spreadingactivation retrieval mechanism that spread acrosslinks within the network, and the more links tra-versed the slower the retrieval time. So, the originalevidence appeared to support the counterintuitiveprediction that subjects indeed needed to go throughthe ‘CANARY IS A BIRD’ link and then the ‘BIRDIS AN ANIMAL’ link to verify that ‘CANARIESBREATHE’, because this is where BREATHES islocated in the network.

The power of network theory to economicallyrepresent the relations among a large amount of infor-mation and the confirmation of the counterintuitivepredictions via the sentence verification studies byCollins and Quillian (1969) clearly encouragedresearchers to investigate the potential of these net-works. However, it soon became clear that the initialhierarchically arranged network structure had somelimitations. For example, the model encounteredsome difficulties handling the systematic differencesin false reaction times, that is, the finding that correct‘false’ responses to ‘BUTTERFLIES ARE BIRDS’ areslower than responses to ‘SPIDERS ARE BIRDS.’Importantly, there was also clear evidence of typical-ity effects within categories. Specifically, categorieshave graded structure, that is, some examples ofBIRD, such as ROBINS, appear to be better examplesthan other BIRDS, such as OSTRICHES.

Canary

Bird

Animal

OstrichCan sing

Is yellow Is tall

Has longthin legs

Has feathers

Has wings Can fly

Has skin Can move around Eats Breathes

SalmonShark

FishHas fins

Is dangerous

Can bite

Has gills

Can swim

Is pink Is edible

Swimsupstream tolay eggs

Can’t fly

Figure 1 Hierarchically arranged network. Taken from Collins A and Quillian MR (1969) Retrieval time from semanticmemory. J. Verb. Learn. Verb. Behav. 8: 240–247.

Semantic Memory 513


There were numerous attempts to preserve thebasic network structure of Collins and Quillian(1969), and indeed, some general models of cognitiveperformance still include aspects of such networkstructure. Collins and Loftus (1975) took a majorstep forward when they developed a network thatwas not forced into a hierarchical framework. This isdisplayed in Figure 2. As shown, these networks arebasically unstructured, with pathways between con-cepts that are related and the strength of therelationship being reflected by the length of the path-ways. Collins and Loftus further proposed that thelinks between nodes could be dependent on semanticsimilarity (e.g., items from the same category, such asDOG and CAT, would be linked), or the links couldemerge from lexical level factors, such as cooccur-rence in the language. Thus, DOG and CAT wouldbe linked because these two items often occur insimilar contexts. Because the strength of spreadingactivation is a function of the distance the activationtraversed, typicality effects can be nicely captured inthis framework by the length of the pathways. Ofcourse, one might be concerned that such networksare not sufficiently constrained by independent evi-dence (i.e., if one is slow the pathway must be long).Nevertheless, such networks have been implemented

to capture knowledge representation in both seman-tic and episodic domains (see Anderson, 2000).

More recently, there has been a resurgent interestin a type of network theory. Interestingly, thesedevelopments are again driven from fields outsideof psychology such as physics (see Albert andBarabasi, 2002) and biology (Jeong et al., 2000).This approach is very principled in nature in that ituses large existing databases to establish the connec-tions across nodes within a network and then usesgraph analytic approaches to provide quantitativeestimates that capture the nature of the networks. Inthis light, researchers are not arbitrarily constructingthe networks but are allowing the known relationsamong items within the network to specify the struc-ture of the network. This approach has been used toquantify such diverse networks as the power grid ofthe Western United States and the neural network ofthe worm, C. elegans (Watts and Strogatz, 1998). Onceone has the network established for a given domain(i.e., providing connections between nodes), one canthen quantify various characteristics of the network,such as the number of nodes, the number of path-ways, the average number of pathways from a node,and the average distance between two nodes.Moreover, there are more sophisticated measuresavailable such as the clustering coefficient, whichreflects the probability that two neighbors of a ran-domly selected node will be neighbors of each other.In this sense, these parameters quantify the charac-teristics of the targeted network. For example, whenlooking at such parameters, Watts and Strogatz(1998) found that naturally occurring networks havea substantially higher clustering coefficient andrelatively short average distances between nodescompared with randomly generated networks thathave the same number of nodes and average connec-tivity between nodes. This general characteristic ofnetworks is called ‘small world’ structure. These highclustering coefficients may reflect ‘hubs’ of connec-tivity and allow one to access vast amounts ofinformation by retrieving information along thehubs. In popular parlance, such hubs may allow oneto capture the six degrees of separation between anytwo individuals that Milgram (1967) proposed andthat has been popularized by the game ‘‘six degrees ofseparation with Kevin Bacon’’.

What do worms, power grids, and parlor gameshave to do with semantic memory? Steyvers andTenenbaum (2005) used three large databases reflect-ing the meaning of words to construct networks ofsemantic memory. These included free-association

StreetVehicle

Car

Ambulance

TruckBus

Fireengine

Red

Fire

House

PearsCherries

ApplesGreen

Orange

Yellow

Roses

Green

Violets

Sunrises

Sunsets

Clouds

Figure 2 Semantic network. From Collins AM and LoftusEF (1975) A spreading-activation theory of semanticprocessing. Psychol. Rev. 82: 407–428.

514 Semantic Memory


norms (Nelson et al., 1998), WordNet (Miller, 1990),and Roget’s Thesaurus (1911). For example, if sub-jects are likely to produce a word in response toanother word in the Nelson et al. free-associationnorms, then a connection between the two nodeswas established in the network. Interestingly,Steyvers and Tenenbaum found that these semanticnetworks exhibited the same small world structure asother naturally occurring networks; specifically, high-clustering coefficients and a relatively small averagepath distance between two nodes. As shown inFigure 3, if one moves along the hub of highly inter-connected nodes, an enormous amount of informationbecomes readily available via traversing a small num-ber of links.

Of course, it is not a coincidence that naturallyoccurring networks have small world structure. Theseductive conclusion here is that knowledge repre-sentation has some systematic similarities acrossdomains. Indeed, Steyvers and Tenenbaum (2005)and others have suggested that such structure reflectscentral principles in development and representationof knowledge. Specifically, Steyvers and Tenenbaumargue that as the network grows, new nodes arepredisposed to attach to existing nodes in a probabil-istic manner. It is indeed quite rare that a newmeaning of a word is acquired without it beingsome variation of a preexisting meaning (see Carey,

1978). Hence, across time, nodes that are added to thenetwork will be preferentially attached to existingnodes. This will give rise to a high degree of localclustering, which is a signature of small world net-work structure. We return to the issue of howconcepts develop in a later section.

It is noteworthy that Steyvers and Tenenbaum(2005) have also provided empirical support fromtheir network analyses. For example, they havefound that word frequency, or the degree to which aword is encountered in language, and age of acquisi-tion, defined as the average age at which a child learnsa given word, effects in naming and lexical decisionperformance naturally fall from this perspective.Naming and lexical decision are two of the mostcommonly used word recognition tasks used inresearch investigating the nature and structure ofsemantic memory. In naming (or speeded pronuncia-tion), a participant is asked to read a presentedstimulus aloud as quickly as possible, whereas inlexical decision, he or she is asked to indicate whethera letter string is a real word or a pseudoword (i.e., astring of letters that does not correspond to the spel-ling of a real word). In both tasks, the primarydependent measure is response latency. The generalassumption is that the speed required to access thepronunciation of a word or to recognize a string ofletters reflects processes involved in accessing stored

ACHE HEAD

PAIN

SOOTHE HURT

EASE

TENSE

CALM

RELAX

COMFORTABLE

RECLINER

SLEEP

REST HULA

VACATION

SUN ISLAND

LAVA

ERUPT

VOLCANO

CRATER

STOMACH

MASSAGE

COMFORT

HAWAII

Figure 3 Segment of small world semantic network. Courtesy of Marc Steyvers.

Semantic Memory 515


knowledge about that word. Interestingly, Steyversand Tenenbaum found a reliable negative correlationbetween number of connections to a node (semanticcentrality) in these networks and response latency,precisely as one might predict, after correlated vari-ables such as word frequency and age of acquisitionhave been partialed out (also see Balota et al., 2004).Clearly, further work is needed to empirically con-firm the utility of these descriptions of semanticstructure and the mechanisms by which such net-works develop over time. However, the recent graphanalytic procedures have taken a significant steptoward capturing semantic memory within an empiri-cally verified network.

2.28.3 Feature Analytic Approaches

An alternative to concepts being embedded within arich network structure is an approach wherein mean-ing is represented as a set of primitive features thatare used in various combinations to represent differ-ent concepts. Of course, this issue (distributedrepresentation of knowledge, by way of features, vs.a localist representation, via a node to concept rela-tionship) is central to attempts to represent andquantify learning and memory in general. We nowturn to a review of the feature-based approaches insemantic memory.

The original Collins and Quillian (1969) researchgenerated a great deal of attention, and soon research-ers realized that categories reflected more gradedstructures than was originally assumed. Specifically,some members of categories are good members(ROBIN for BIRD), whereas other members appearto be relatively poor members (VULTURE forBIRD) but are still definitely members of the category(see Battig and Montague, 1969; Rosch, 1973). Inaddition, there was a clear influence of goodness ofan exemplar on response latencies in the sentenceverification task described above. Specifically, goodexemplars were faster to verify than poor exemplars,referred to as the typicality effect. The Collins andQuillian hierarchical network model did not have anyobvious way of accommodating such degrees of cate-gory membership.

Smith et al. (1974) took a quite different approachto accommodate the results from the sentence ver-ification task. They rejected the strong assumptionsof network theory and proposed a model that empha-sized the notion of critical semantic features inrepresenting the meaning of a word. So, for example,

the word BIRD might be represented as animal, twolegged, has wings, sings, is small, flies, and so on.There is no hierarchical organization within thismodel, but concepts reflect lists of critical features.They also distinguished between two classes of fea-tures, defining features and characteristic features.Defining features are the necessary features that anexemplar must have to be a member of a category. So,for example, all birds must eat, move, lay eggs, and soforth. On the other hand, characteristic features arefeatures that most, but not all, exemplars have, suchas small, flies, sings.

The second important aspect of the Smith et al.(1974) perspective is the emphasis on the decisionprocesses engaged in the classic sentence verifica-tion task (see Atkinson and Juola, 1974; Balota andChumbley, 1984, for similar decision models appliedto short-term memory search and lexical decision,respectively). In verifying a sentence such as ‘AROBIN IS A BIRD,’ subjects first access all (bothdefining and characteristic) features associated withROBIN and all features associated with BIRD. Ifthere is a high degree of overlap in the features,that is, above some criterion, the subject can make afast ‘yes’ response. This would be the case in ‘AROBIN IS A BIRD,’ since both defining and char-acteristic features provide a high degree of overlap.On the other hand, some exemplars of a given cate-gory may overlap less in characteristic features suchas in ‘AN OSTRICH IS A BIRD’. Although ostrichesare clearly birds, they are not small and do not fly,which are characteristic features of birds. Hence, insuch cases, the subject needs to engage in an addi-tional analytic checking process in which only thedefining features are compared. This additionalcheck process takes time and so slows response laten-cies. Hence, the model can naturally capture thetypicality effects mentioned above, that is, robinsare better exemplars of birds than ostriches, becauserobins can be verified based on global overlap infeatures, whereas ostriches must engage the second,more analytic comparison of only the defining fea-tures, thereby slowing response latency.

In addition to accounting for typicality effects, thefeature analytic model also captured interesting dif-ferences in latencies to respond ‘no’ in the sentenceverification task. Specifically, subjects are relativelyfast to reject ‘A CARP IS A BIRD’ compared with ‘ABUTTERFLY IS A BIRD.’ Carps do not have manyoverlapping features with birds, and so the subjectcan quickly reject this item, that is, there is virtuallyno overlap in features. However, both butterflies and

516 Semantic Memory


birds typically have wings, are small, and fly. Hence,the subject must engage the additional check of thedefining features for ‘BUTTERFLY IS A BIRD,’which ultimately leads to slower response latencies,compared with the sentence ‘A CARP IS A BIRD.’

Although there were clear successes of the Smithet al. (1974) feature analytic approach, there werealso some problems. For example, the model wascriticized for the strong distinction between charac-teristic and defining features. In fact, McCloskey andGlucksberg (1979) provided a single process randomwalk model that accommodated many of the sameresults of the original Smith et al. model withoutpostulating a distinction between characteristic anddefining features. According to the random walkframework, individuals sample information acrosstime that supports either a yes or no decision. If thefeatures from the subject and predicate match, thenmovement toward the yes criterion takes place; if thefeatures do not match, then movement toward the nocriterion takes place. This model simply assumedthat the likelihood of sampling matching featureinformation for the subject and predicate is greaterfor typical members than nontypical members, andtherefore the response criterion is reached morequickly for typical than nontypical members, thusproducing the influence on response latencies. Thedistinction between single- and dual-process modelsis a central issue that pervades much of cognitivescience.

A second concern about the Smith et al. (1974)model is that they did not directly measure featuresbut, rather, inferred overlap in features based onmultidimensional scaling techniques, in which anindependent group of subjects simply rated the simi-larity of words used in the sentence verificationexperiment. In this way, one could look at the simi-larity of the words along an N-dimensional space.Interestingly, Osgood et al. (1957) used a similarprocedure to tackle the meaning of words in theirclassic work on the semantic differential. Osgoodet al. found that when subjects rated the similarityacross words, and these similarity ratings were sub-mitted to multidimensional scaling procedures, therewere three major factors that emerged: Evaluative(good–bad), potency (strong–weak), and activity(active–passive). Although clearly this work is pro-vocative, such similarity ratings do not provide adirect measure of the features available for a concept.So, if there are indeed primitive features, it seemsnecessary to attempt to more directly quantify thesefeatures.

McRae and colleagues have been recentlyattempting to provide such constraints on featureanalytic models (McRae et al., 1997; Cree et al.,1999; Cree and McRae, 2003; McRae, 2004; McRaeet al., 2005). The goal here is to develop a feature-based computational model implemented in anattractor network capable of capturing the statisticalregularities present in semantic domains. The gen-eral notion underlying attractor network models isthat knowledge is distributed across units (whichmight be thought of as features) and that the networksettles into a steady pattern of activity that reflectsthe representation of a concept. The conceptualrepresentations that form the basis of semanticknowledge in the model are derived from featurenorming data. To collect norms, groups of partici-pants are asked to list features for a number ofconcepts (e.g., for DOG, participants might listBARKS, FURRY, CHASES CATS, etc.). McRae andcolleagues propose that when participants are asked tolist features of various basic-level category exemplars(e.g., DOG and APPLE are basic-level concepts fromthe superordinate category of MAMMALS andFRUIT, respectively), the resulting lists of featuresreflect the explicit knowledge people have of theseconcepts. Importantly, McRae does not claim that thenature of the representation consists of a feature list;rather, he argues, the features are derived fromrepeated multisensory interactions with exemplars ofthe concept, and in a feature listing task, subjectstemporarily create an abstraction for the purposeof listing features that can be verbally described.Currently, feature norming data are available for 541concrete objects, representing a wide variety of basic-level concepts. Importantly, the model can account formany empirical observations in semantic tasks, as dis-cussed below.

The major assumption implemented by McRaeand colleagues’ model is that semantic knowledge,as represented by feature lists, involves the statisticalaveraging of feature correlations among members ofsimilar categories. Features are correlated if they co-occur in basic-level concepts. For example, HASFUR is highly correlated with HAS FOUR LEGS,as these two features cooccur in numerous exemplarsof the mammal category. However, HAS FUR andHAS WINGS have a low (almost nonexistent) cor-relation, as these two features do not co-occurfrequently. The argument is that individuals arehighly sensitive to the regularity of the correlations,which are tapped by semantic tasks. As demonstratedby McRae et al. (1999), the strength of the feature

Semantic Memory 517


correlations predicted feature verification latencies inboth human subjects and model simulations, withstronger correlations yielding faster response latenciesthan weaker correlations when the concept name waspresented before the feature name (e.g., DOG-FUR).In addition, the correlation strength interacted withstimulus onset asynchrony (the time between theonset of the concept name and the onset of the featurename, SOA). Specifically, the effect of feature correla-tions was larger at shorter SOAs, with only highcorrelations predicting response latencies. However,at longer SOAs, even weakly correlated features influ-enced response times, indicating that, as more timewas allowed for the effects of correlated features toemerge, even the more weakly correlated feature-concept pairs benefited from the shared representa-tion. In another series of studies, McRae et al. (1997)reported that the strength of feature correlations pre-dicted priming for exemplars from the living thingsdomain but not for exemplars from nonliving thingsdomains, for which priming was instead predicted byindividual features. This finding is consistent withevidence that, compared to living things, nonlivingthings tend to have a lower degree of correlatedfeatures (also see section 2.28.8.1).

Several interesting extensions of McRae and col-leagues’ work on the role of features in organizingsemantic knowledge have been recently reported.Pexman et al. (2002) examined the role of the numberof features associated with a concept and found thatitems with more features were responded to faster inboth naming and lexical decision tasks after a numberof other variables known to influence visual wordrecognition latencies had been factored out. Pexmanet al. (2003) reported similar results in a semanticcategorization task and in a reading task. These find-ings were interpreted as supporting the distributednature of semantic representations in which featuresare assumed to reflect access to conceptual knowl-edge, and this information quickly comes on line inisolated word recognition tasks.

In a related vein, there is recent evidence from thecategorization literature that categories with richerdimensionalities (i.e., more features and more corre-lations among features) are easier to learn thancategories with fewer dimensions (Hoffman andMurphy, 2006). Thus, rather than resulting in com-binatorial explosions that make learning impossible,rich categories with many features lend themselveswell to learning – a finding that is nicely mirrored inhow people, even very young children, quickly andreliably learn to recognize and classify objects in the

world. Indeed, it seems that learning to categorizecomplex objects, which might be quite similar interms of features, is something most individuals cando reliably and easily. One concern that arises whenone examines the richness of the stimuli in the envi-ronment, is the potentially infinite number offeatures that are available to identify a given concept.In fact, critics of feature-based models have arguedthat the number of possible feature combinationswould result in combinatorial explosion, as knowingeven a few features of a category could easily result inan enormous number of ways in which the featurescould be correlated and integrated (see Murphy,2002, for a discussion). However, as McRae (2004)notes, two points are relevant in addressing this issue.The first is that the feature correlations tend toinfluence performance largely in implicit tasks –thus reducing the necessity of explaining how anindividual can explicitly use the vast amount ofinformation available. The second point is that thefeature vectors that underlie semantic representa-tions are generally sparse. In other words, theabsence of a specific feature is uninformative, so, forexample, knowing that a dog does not have feathers isrelatively uninformative. Thus, although feature-based models might not fully capture the richnessof the knowledge that individuals have about con-cepts, they have been useful in advancing research inthe field of semantics.

2.28.4 Concept Learning andCategorization

Since semantic memory deals with the nature ofrepresentation of meaning, and categories are centralto meaning, it is important to at least touch on thearea of categorization and how concepts develop. Intheir classic book, Bruner et al. (1956) emphasized theimportance of categorization in organizing whatappears to be a limitless database that drives complexhuman learning and thought. Categorization hasbeen viewed as a fundamental aspect of learningand indeed has been observed early in childhood(Gelman and Markman, 1986) and in other speciessuch as pigeons (Herrnstein et al., 1976). Ross (SeeChapter 2.29) provides a much more focused discus-sion of this topic.

One intriguing question that arises when one con-siders the content of semantic memory concerns thegrain size and structure of the representations. Inother words, is there a level at which objects in the

518 Semantic Memory


real world are more or less easy to learn and categor-ize? One possibility is that the world is initiallyperceived as a continuum in which there are notseparate ‘things.’ Through repeated interactionswith verbal labels or other forms of learning, anindividual learns how to discriminate separate objects(e.g., Leach, 1964). This approach places the burdenon an extensive and demanding learning process. Analternative approach is that the human cognitivesystem is ideally suited to detect and recognizeobjects at a specific grain or level. The assumptionthat the system is biased toward recognizing specificpatterns implies that the process of learning theappropriate verbal labels that refer to specific itemsin the environment is significantly easier. This prob-lem – how very young children learn that when theirmother points to a dog and says ‘dog’ the referent ofthe phonological pattern in question refers to anentire object, and not to furry things, things of acertain color, or loosely attached dog parts – hasbeen extensively discussed by Quine (1960).

In an elegant series of experiments, Rosch andcolleagues (e.g., Rosch et al., 1976) provided empiricalevidence that there is indeed a specific level at whichcategories of objects are represented that contains themost useful amount of information. For example,identifying a given object as a DOG implies that onerecognizes that the specific exemplar is a dog,although it may differ from other dogs one hasencountered. Simply knowing something is a dogallows one to draw upon a pool of stored knowledgeand experiences to infer appropriate behaviors andinteractions with the categorized object. However,knowing the object is an animal is not as informative,given the wide variability among animals. For exam-ple, interactions with an elephant are likely to be quitedifferent from those one might have with a spider.Conversely, classifying the exemplar as a Collie or asa German Shepherd does not add a significant amountof inferential power for most purposes.

Rosch et al. (1976) argued that at the basic level,categories are highly informative and can be reliablyand easily discriminated from other categories.Exemplars of basic-level categories (e.g., DOGS,BIRDS, CARS, etc.) have many attributes in com-mon, tend to be similar in shape and in how oneinteracts with them, and allow easy extraction of aprototype or summary representation. The prototypecan be accessed and serves as a benchmark againstwhich novel exemplars can be compared: Those thatare highly similar to the prototype will be quicklyand easily classified as members of the category.

Exemplars that differ from the prototype will berecognized as less typical members of a category(e.g., penguins are quite different from many otherbirds). Hence, typicality effects fall quite nicely fromthis perspective.

Historically, there has again been some tensionbetween abstract prototype representation and morefeature-based approaches. Consider the classic workby Posner and Keele (1968). Although cautious intheir interpretation, these researchers reported evi-dence suggesting that prototypes (in this case acentral tendency of dot patterns) were naturallyabstracted from stored distortions of that prototype,even though the prototype was never presented forstudy. They also found that variability across a suffi-cient number of distortions was critical for abstractingthe prototype. These results would appear to supportthe notion that there is a natural tendency to abstractsome representation that is a central tendency ofexemplars that share some common elements. So,‘dogness’ may be abstracted from the examples ofdogs that one encounters. This could suggest thatthere is indeed a unified representation for dogness.

An alternative approach is that there is nothingunique about these central tendencies but, rather,such representations reflect the similarity of the epi-sodically stored representations in memory. This is aparticularly important observation because it sug-gests that there is a blending of different types ofmemories, that is, categorical information is simplydecontextualized episodic memories. Consider, forexample, the classic MINERVA model developedby Hintzman (1986, 1988). In this computationalmodel, each episodic experience lays down a uniquetrace in memory, which is reflected by a vector oftheoretical features. There is no special status ofcategory representations or hierarchical structure.Rather, categorization occurs during retrieval whena probe (the test item) is presented to the system, andthe feature vector in the probe stimulus is correlatedwith all the episodically stored traces. The familiarityof a test probe is a reflection of the strength of thecorrelations among elements in memory. Because theschema overlaps more with multiple stored represen-tations, that is, it is the central tendency, it willproduce a relatively high familiarity signal orstrength in a cued recall situation. The importanceof the Hintzman approach is that there is no need todirectly store central tendencies, as they naturallyarise out of the correlation among similar storedtraces in the feature vectors. Moreover, asHintzman argues, there is no need to propose a

Semantic Memory 519


qualitative distinction between episodic and semanticmemories, because both rely on the same memorysystem, that is, a vast storehouse of individual fea-ture-based episodic traces.

The notion that categories are a reflection ofsimilarity structure across memory traces and canbe generated during retrieval clearly has someappeal. Indeed, Barsalou (1985) demonstrated theimportance of ad hoc categories that seem to be easilygenerated from traces that do not inherently havenatural category structure; for example, what dophotographs, money, children, and pets have in com-mon? On the surface, these items do not appear to besimilar – they do not belong to the same taxonomiccategory, nor do they share many features. However,when given the category label ‘‘things to take out ofthe house in the case of a fire,’’ these items seem to fitquite naturally together because our knowledge basecan be easily searched for items that are in the houseand are important to us. As Medin (1989) has argued,similarity depends on the theoretical frame that aparticipant uses to guide a search of memory struc-tures. There appears to be an unlimited number ofways in which similarity can be defined, and hencesimilarity discovered. For example, lichen and squir-rels are similar if one is interested in specifying thingsin a forest. This brings us to the remarkable contextdependency of meaning, and the possibility thatmeaning is not defined by the stimulus per se but isa larger unit involving both the stimulus and thesurrounding context. The word DOG in the contextof thinking about house pets compared with the wordDOG in the context of guard dogs or drug-sniffingdogs probably access quite different interpretations,one in which the focus is on companionship, furri-ness, and wagging tails, the other in which the morethreatening aspects of dogness, such as sharp teeth,are accessed. One might argue that the context acti-vates the relevant set of features, but even this isdifficult until one has sufficient constraint on whatthose features actually are.

2.28.5 Grounding Semantics

In part because of the difficulties in defining thecritical features used to represent meaning andpotential problems with the tractability of prototypesof meaning, several researchers attempted to takenovel approaches to the nature of the representations.There are two general approaches that we review inthis section. First, because of the increase in

computational power, there has been an increasedreliance on analyses of large-scale databases to extractsimilarities across the contexts of words used in var-ious situations. This perspective has some similarity tothe exemplar-based approach proposed by Hintzman(1986, 1988) and others described earlier. In this sense,meaning is grounded in the context in which wordsand objects appear. The second approach is to con-sider the perceptual motor constraints afforded byhumans to help ground semantics, that is, the embo-died cognition approach. We review each of thesein turn.

2.28.5.1 Grounding Semantics in Analysesof Large-Scale Databases

This approach attempts to directly tackle the povertyof the stimulus problem when considering the knowl-edge that humans have acquired. Indeed, since thedays of Plato, philosophers (and more recently psy-chologists and linguists) have attempted to resolvethe paradox of how humans can acquire so muchinformation based on so little input. Specifically,how is it that children learn so much about thereferents of words, when to use them, what theirsyntactic class is, what the relations among referentsare, and so on, without explicit instruction? Somehave argued (e.g., Pinker, 1994) that the poverty ofthe stimulus is indeed the reason one needs to buildin genetically predisposed language acquisitiondevices. However, recent approaches to this issue(e.g., Latent Semantic Analysis, or LSA, Landauerand Dumais, 1997; Hyperspace Analogue toLanguage, or HAL, Burgess and Lund, 1997) havesuggested that the stimulus input is not so impover-ished as originally assumed. One simply needs morepowerful statistical tools to uncover the underlyingmeaning and the appropriate database.

In an attempt to better understand how rich thestimulus is when embedded in context, Landauer andDumais (1997) analyzed large corpora of text thatincluded over 4.6 million words taken from anEnglish encyclopedia, a work intended for youngstudents. This encyclopedia included about 30 000paragraphs reflecting distinct topics. From this, theauthors constructed a data matrix that basicallyincluded the 60 000 words across the 30 000 para-graphs. Each cell within the matrix reflected thefrequency that a given word appeared in a givenparagraph. The data matrix was then submitted to asingular value decomposition, which has strong sim-ilarities to factor analysis to reduce the data matrix to

520 Semantic Memory


a limited set of dimensions. Essentially, singularvalue decomposition extracts a parsimonious repre-sentation of the intercorrelations of variables, but,unlike factor analysis, it can be used with matricesof arbitrary shape in which rows and columns repre-sent the words and the contexts in which the wordsappear. In this case, the authors reduced the matrix to300 dimensions. These dimensions reflect the inter-correlations that arise across the words from thedifferent texts. So, in some sense the 300 dimensionsof a given word will provide information about thesimilarity to all other words along these 300 dimen-sions, that is, the degree to which words co-occur indifferent contexts. The exciting aspect from this datareduction technique is that by using similarity esti-mates, the model actually performs quite well incapturing the performance of children acquiring lan-guage and adults’ performance on tests based onintroductory textbooks. In this way, the meaning ofa word is being captured by all the past experienceswith the word, the contexts with which that word(neighbors) occurs, the contexts that the neighborsoccur in, and so on.

The remarkable success of LSA, and other similarapproaches such as HAL (Burgess and Lund, 1997),provides a possible answer to the poverty of thestimulus problem, that is, when considering the con-text, the stimulus is indeed very rich. In the past, wesimply have not been able to analyze it appropriately.Moreover, the model nicely captures the apparentcontextual specificity of meaning in that meaning isdefined by all the contexts that words have appearedin and hence will also be constantly changing ever soslightly across subsequent encounters. Finally, themodel is indeed quite important because it does notrely on a strong distinction between semantic andepisodic memory since it simply reflects past accu-mulated exposure to language. In this sense, it hassome similarity to the Hintzman (1986) modeldescribed above.

2.28.5.2 Grounding Semantics inPerceptual Motor Systems

An alternative approach that has been receiving con-siderable recent attention is that meaning can begrounded in perceptual-motor systems (e.g.,Barsalou, 1999). Briefly, this perspective is part of theembodied cognition approach that posits that the cog-nitive system of any organism is constrained by thebody in which it is embedded (Wilson, 2002). Thus,cognition (in this case meaning) is not viewed as being

separable from perceptual, motor, and proprioceptivesystems; rather, it is through the interactions of thesesystems with the environment that cognition emerges.Furthermore, the type of representations that anorganism will develop depends on the structure ofthe organism itself and how it exists in the world.This approach has its roots in Gibson’s (1979) ecolo-gical psychology, as it is assumed that structures in theenvironment afford different interactions to differentorganisms. It is through repeated interactions withthe world that concepts and knowledge emerge.Importantly, the very nature of this knowledge retainsits connections to the manner in which it was acquired:Rather than assuming that semantic memory consistsof amodal, abstract representations, proponents ofembodied approaches argue that representations aregrounded in the same systems that permitted theiracquisition in the first place (Barsalou et al., 2003).

According to the modality-specific approaches toknowledge, a given concept is stored in adjacent mem-ory systems rather than being abstracted. For example,in Barsolou’s (1999) account, knowledge is stored inperceptual symbol systems that emerge throughrepeated experience interacting with an object or anevent. Briefly, Barsalou assumed that when a perceptis encountered, selective attention focuses on context-relevant aspects of the percept and allows modalrepresentations to be stored in memory. Repeatedinteractions with similar events or members of thesame category result in the formation of a complex,multimodal representation, and a simulator emergesfrom these common representations. Simulators arethe basic unit of the conceptual system and consist ofa frame (which is somewhat similar to a schema), thepurpose of which is to integrate the perceptual repre-sentations. Simulators provide continuity in thesystem. Importantly, the representations that arestored include not only modal, perceptual information(e.g., sounds, images, physical characteristics) but alsoemotional responses, introspective states, and proprio-ceptive information. Retrieving an exemplar orremembering an event is accomplished by engagingin top-down processing and activating the targetedsimulator. Importantly, a given simulator can yieldmultiple simulations, depending on the organism’sgoal, the context, and the relevant task demands. Forexample, different simulations for DOG are possible,such that a different pattern of activity will occur if thewarm and furry aspect of dog is relevant or whetherthe aspect of being a guard or police dog is relevant.Of course, this nicely captures the context sensitivityof meaning. Barsalou (1999) argues that perceptual

Semantic Memory 521


symbol systems are as powerful and flexible as amodalmodels, as they are able to implement a completeconceptual system (see also Glenberg and Robertson,2000).

Evidence in support of modal approaches tosemantics can be found in both behavioral and cog-nitive neuroscience studies. We briefly review someof this evidence here, although a full review of theneuroscience literature is beyond the scope of thischapter (See Chapter 3.07 for further discussion ofthis area). For example, there is evidence from lesionstudies that damage to the pathways supporting aspecific modality results in impaired performance incategorization and conceptual tasks that rely on thatsame modality. Specifically, damage to visual path-ways generally results in greater impairment in thedomain of living things, which tend to rely heavily onvisual processes for recognition. Conversely, damageto motor pathways tends to impair knowledge ofartifacts and tools, as the primary mode of interactionwith these items is through manipulation (see Martin,2005). Consistent with the lesion data, neuroimagingstudies indicate that different regions of the cortexbecome active when people process different cate-gories. Regions adjacent to primary visual areasbecome active when categories such as animals areprocessed (even if the presentation of the stimulusitself is not in the visual modality), whereas regionsclose to motor areas become active when categoriessuch as tools are processed. These findings have beeninterpreted as consistent with the hypothesis thatpeople run perceptual-motor simulations when pro-cessing conceptual information (Barsalou, 2003).

Pecher et al. (2003) reported evidence from aproperty verification task indicating that participantswere faster in verifying properties in a given modality(e.g., BLENDER-loud) after verifying a differentproperty for a different concept in the same modality(e.g., LEAVES-rustling) than when a modality switchwas required (e.g., CRANBERRIES-tart). Pecheret al. argued that the switch cost observed was con-sistent with the hypothesis that participants ranperceptual simulations to verify the properties (inthis case sounds) rather than accessing an amodalsemantic system. In a subsequent study, Pecher et al.(2004) observed that when the same concept in aproperty verification task was paired with two prop-erties from different modalities, errors and latenciesincreased when verifying the second property. Pecheret al. interpreted this finding as indicating that recentexperiences with a concept influence the simulationof the concept. Importantly, researchers have argued

that such results are not simply a result of associativestrength (i.e., priming) nor of participants engaging inintentional imagery instructions (Barsalou, 2003;Solomon and Barsalou, 2004).

Although the results summarized above are com-pelling and are supportive of the hypothesis thatsensory-motor simulations underlie many semantictasks, the majority of these studies have examinedtasks such as property verification and property gen-eration. The question thus arises of whether theresults are somehow an artifact of the task demands,and specifically whether these results reflect thestructure of the semantic memory system or whethersubjects are explicitly retrieving information as theynotice the relations embedded within the experimen-tal context. Glenberg and Kaschak (2002) extendedthe evidence for embodiment effects to a novel seriesof tasks that do not appear as susceptible to taskdemand effects. In these experiments, participantsread a brief sentence and judged whether the sen-tence made sense or not. The critical sentencescontained statements that implied motion eithertoward the participant (e.g., ‘‘Nancy gave you thebook’’) or away from the participant (e.g., ‘‘You gavethe book to Nancy’’). Participants responded by mov-ing their hand toward themselves or away fromthemselves. Glenberg and Kaschak found what theycalled the action-sentence compatibility effect: Whenthe required response was consistent with the move-ment implied in the sentence, participants were fasterthan when the implied motion and the actual physi-cal response were inconsistent. These data appearmost consistent with the view that when processinglanguage, people relate the meaning of the linguisticstimulus to action patterns.

2.28.6 Measuring SemanticRepresentations and Processes:Insights from Semantic PrimingStudies

As described above, there have been many empiricaltools that have been used to provide insights into thenature of semantic memory. For example, as notedearlier, some of the early work by Osgood et al.(1957) attempted to provide leverage on fundamentalaspects of meaning via untimed ratings of large sets ofwords and multidimensional scaling techniques.With the advent of interest in response latencies,researchers turned to sentence verification tasksthat dominated much of the early work in the 1970s

522 Semantic Memory


and 1980s. Although this work has clearly been influ-ential, the explicit demands of such tasks (e.g.,explicitly asking subjects to verify the meaningful-ness of subject-predicate relations) led someresearchers in search of alternative ways to measureboth structure and retrieval processes from semanticmemory. There was accumulating interest in auto-matic processes (LaBerge and Samuels, 1974; Posnerand Snyder, 1975) that presumably captured themodular architecture of the human processing sys-tem (Fodor, 1983), and there was an emphasis onindirect measures of structure and process. Hence,researchers turned to semantic priming paradigms.

Meyer and Schvaneveldt (1971) are typicallyregarded as reporting the first semantic primingstudy. In this study, subjects were asked to make lexicaldecisions (word-nonword decisions) to pairs of stimuli.The subjects’ task was to respond yes only if bothstrings were words. The interesting finding here wasthat subjects were faster to respond yes when the wordswere semantically related (DOCTOR NURSE), com-pared with when they were unrelated (BREADNURSE). This pattern was quite intriguing becausesubjects did not need to access the semantic relationbetween the two words to make the word/nonworddecisions. Hence, this may reflect a relatively puremeasure of the underlying structure and retrievalprocesses, uncontaminated by explicit task demands.Moreover, the development of this paradigm was quiteimportant because researchers thought it may tap thespreading activation processes that was so central totheoretical developments at the time.

The research on semantic priming took a signifi-cant leap forward with the dissertation work of Neely(1977), who used a framework developed by Posnerand Snyder (1975) to decouple the attentional strat-egic use of the prime-target relations from a moreautomatic component. In this study, subjects onlymade lexical decisions to the target, and subjectswere given explicit instructions of how to use theprime information. For example, in one condition,subjects were told that when they received theprime BODY, they should think of building parts(Shift condition), whereas in a different condition,subjects were told that when they received the cate-gory prime BIRD, they should think of birds(Nonshift condition). Neely varied the time availableto process the prime before the target was presentedby using SOAs ranging from 250 to 2000ms. Theimportant finding here is that the instructions ofwhat to expect had no influence on the primingeffects at the short SOA (i.e., priming occurred if the

prime and target had a semantic relationship, inde-pendent of expectancies), but they did have a largeeffect at the long SOA, when subjects had time toengage an attentional mechanism (i.e., the primingeffects were totally dependent on what subjects weretold to expect, independent of the preexisting rela-tionship). Hence, Neely argued that the short SOAdata reflected pure automatic measures of the seman-tic structure and retrieval processes and could be usedas a paradigm to exploit the nature of such semanticrepresentations.

A full review of the rich semantic priming litera-ture is clearly beyond the scope of the presentchapter (see Neely, 1991; Lucas, 2000; Hutchison,2003, for excellent discussions of the methodologicaland theoretical frameworks). However, it is useful tohighlight a few issues that have been particularlyrelevant to the current discussion. First, there issome controversy regarding the types of prime-targetrelations that produce priming effects. For example,returning to the initial observation by Meyer andSchvaneveldt (1971), one might ask if DOCTORand NURSE are related because they share someprimitive semantic features or are simply relatedbecause they are likely to co-occur in the same con-texts in the language. Of course, this distinctionreflects back on core assumptions regarding the na-ture of semantic information, since models like LSAmight capture priming between DOCTOR andNURSE, simply because the two words are likely tocooccur in common contexts. Researchers haveattempted to address this by selecting items thatvary on only one dimension (see, e.g., Fischler,1977; Lupker, 1984; Thompson-Schill et al., 1998).Here, semantics is most typically defined by categorymembership (e.g., DOG and CAT are both semanti-cally related and associatively related, whereasMOUSE and CHEESE are only associativelyrelated). Hines et al. (1986), De Mornay Davies(1998), and Thompson-Schill et al. (1998) have allargued that priming is caused by semantic featureoverlap because of results indicating priming only forwords that shared semantic overlap versus those didnot, when associative strength was controlled.However, Hutchison (2003) has recently arguedthat the studies that have provided evidence forpure semantic effects (i.e., while equating for associa-tive strength), actually have not adequatelycontrolled for associative strength based on theNelson et al. (1998) free-association norms. Clearly,equating items on one dimension (associativestrength or semantic overlap) while manipulating

Semantic Memory 523


the other dimension is more difficult than initiallyassumed. In this light, it is interesting to note thattwo recent review papers have come to different con-clusions regarding the role of semantics in semanticpriming based on such item selection studies. Lucas(2000) argued that there was clear evidence of puresemantic effects, as opposed to associative effects,whereas Hutchison (2003) was relatively more skep-tical about the conclusions from the availableliterature.

Balota and Paul (1996) took a different approach tothe meaning versus associative influence in primingvia a study of multiple primes, instantiating the con-ditions displayed in Table 1. As one can see, theprimes were either both related, first related, secondrelated, or both unrelated to the targets, and the targetscould either be homographic words with distinctmeanings (e.g., ORGAN) or a nonhomographic words(e.g., STRIPES). As one can see, the primes wererelated to the targets at both the semantic and asso-ciative level for the nonhomographs (e.g., LION andSTRIPES are both related to TIGER at the associa-tive and semantic level), but for the homographs theprimes were related to the targets at only the asso-ciative level (e.g., PIANO and KIDNEY are onlyrelated to ORGAN at the associative level, sinceKIDNEY and PIANO are different meanings ofORGAN). Thus, one could compare priming effectsin conditions in which primes converged on the samemeaning of the target (nonhomographs) and primingeffects where the primes diverged on different

meanings (homographs). The results from fourexperiments indicated that the primes produced clearadditive effects, that is, priming effects from the singlerelated prime conditions nicely summated to predictthe priming effects from the double related primeconditions for both homographs and nonhomographs,suggesting that the effects were most likely a result ofassociative level information. Only when subjectsdirected attention to the meaning of the word, viaspeeded semantic decisions, was there any evidenceof the predicted difference between the two conditions.Hence, these results seem to be supportive of thenotion that standard semantic priming effects are likelyto be the result of associative-level connections insteadof meaning-based semantic information. Of course, theinteresting theoretical question is how much of oursemantic knowledge typically used is caused by over-lap in the contexts in which items are stored as opposedto abstracted rich semantic representations.

Hutchison (2003) notes two further findings thatwould appear to be supportive of associative influ-ences underlying semantic priming effects. First, onecan find evidence of episodic priming in lexical deci-sion and speeded word naming tasks. In these studies,subjects study unrelated words such as (CITY-GRASS) and are later presented prime-target pairsin a standard lexical decision study. The interestingfinding here is that one can obtain priming effects insuch studies, compared to an unrelated/unstudiedpair of words (see McKoon and Ratcliff, 1979).Thus, the semantic priming effects obtained inword recognition tasks can also be produced viapurely associative information that develops withina single study exposure. However, it should be notedhere that there is some question regarding the locusof such priming effects and that one needs to beespecially cautious in making inferences from theepisodic priming paradigm and the role of task-specific strategic operations (see, e.g., Neely andDurgunoglu, 1985; Durgunoglu and Neely, 1987;Spieler and Balota, 1996; Pecher and Raajmakers,1999; Faust et al., 2001).

The second pattern of results that Hutchison(2003) notes as being critical to the associativeaccount of semantic priming effects is mediatedpriming. In these situations, the prime (LION) isrelated to the target (STRIPES) only through a non-presented mediator (TIGER). So, the question iswhether one can obtain priming from LION toSTRIPES, even though these two words appear tobe semantically unrelated. Although de Groot (1983)failed to obtain mediated priming effects in the

Table 1 Prime-target conditions from the Balota andPaul (1996) multiprime study

Nonhomographs

Condition Prime 1 Prime 2 Target

Related–related LION TIGER STRIPESUnrelated–related FUEL TIGER STRIPESRelated–unrelated LION SHUTTER STRIPESUnrelated–unrelated FUEL SHUTTER STRIPES

Homographs

Condition Prime 1 Prime 2 Target

Related–related KIDNEY PIANO ORGANUnrelated–related WAGON PIANO ORGANRelated–unrelated KIDNEY SODA ORGANUnrelated–unrelated WAGON SODA ORGAN

Balota DA and Paul ST (1996) Summation of activation: Evidencefrom multiple primes that converge and diverge within semanticmemory. J. Exp. Psychol. Learn. Mem. Cogn. 22: 827–845.

524 Semantic Memory


lexical decision task, Balota and Lorch (1986) arguedthat this may have resulted from the task-specificcharacteristics of this task. Hence, Balota and Lorchused a speeded pronunciation task and foundclear evidence of mediated priming. Evidence formediated priming has also now been found in versionsof the lexical decision task designed to minimize task-specific operations (e.g., McNamara and Altarriba,1988; McKoon and Ratcliff, 1992; Sayette et al.,1996; Livesay and Burgess, 1998). Of course, it isunclear what semantic features overlap betweenLION and STRIPES, and so these results wouldappear to be more consistent with an associative net-work model, in which there is a relationship betweenLION and TIGER and between TIGER andSTRIPES, along with a spreading activation retrievalmechanism (see McKoon and Ratcliff, 1992; Chwillaand Kolk, 2002, for alternative accounts of the retriev-al mechanism).

In sum, although the semantic priming paradigmhas been critical in measuring retrieval mechanismsfrom memory, the argument that these effects reflectamodal semantic representations that are distinctfrom associative information has some difficultyaccommodating the results from multiprime studies,episodic priming studies, and mediated priming stud-ies. As noted earlier, there are available models ofsemantic memory (e.g., Burgess and Lund, 1997,HAL; Landauer and Dumais, 1997, LSA) and cate-gorization (e.g., Hintzman, 1996, MINERVA) thatwould strongly support the associative contributionsto performance in such tasks and, indeed, questionthe strong distinction between semantic and episodicmemory systems. Hence, this perspective predicts astrong interplay between the systems. We now turnto a brief discussion of the evidence that directlyaddresses such an interplay.

2.28.7 The Interplay BetweenSemantics and Episodic Memory

Memory researchers have long understood the influ-ence of preexisting meaning on learning and memoryperformance (see Crowder, 1976, for a review). Indeed,in his original memory manifesto, Ebbinghaus (1885)was quite worried about this influence and so purpose-fully stripped away meaning from the to-be-learnedmaterials by presenting meaningless trigrams (KOL)for acquisition. Of course, semantics has penetratedepisodic memory research in measures of categoryclustering (see Bousfield, 1953; Cofer et al., 1966;

Bruce and Fagan, 1970), retrieval-induced inhibition(see Anderson et al., 1994), and release from proactiveinterference (see Wickens, 1973), among many otherparadigms. Indeed, the interplay between preexistingknowledge and recall performance was the centerpieceof the classic work by Bartlett (1932). Researchersrealized that even consonant–vowel–consonant tri-grams were not meaningless (see Hoffman et al.,1987). At this level, one might even question what itwould mean to episodically store in memory totallymeaningless information.

One place where researchers have attempted tolook at the interplay between semantic and episodicstructures is within the episodic priming paradigmdescribed earlier. In these studies, participantsreceive pairs of unrelated words for study and thenare later given prime-target pairs that have eitherbeen paired together or not during the earlier acqui-sition phase. For example, Neely and Durgunoglu(1985) investigated the influence of studying previouspairs of words and word–nonword combinations onboth lexical decision performance and episodic recog-nition performance (also see Durgunoglu and Neely,1987). Although there were clear differences betweenthe tasks in the pattern of priming effects (suggestingdissociable effects across the two systems), there werealso some intriguing similarities. For example, therewas evidence of inhibition at a short prime-targetSOA (150ms) in both the episodic recognition taskand the lexical decision task from semantically relatedprimes that were in the initial studied list but were notpaired with the target. It appeared as if this additionalsemantic association had to be suppressed in order forsubjects to make both the episodic recognition deci-sion and the lexical decision. The finding that thiseffect occurred at the short SOA also suggests that itmay have been outside the attentional control of theparticipant.

The power of preexisting semantic represen-tations on episodic tasks has recently taken asubstantial leap forward with the publication of animportant paper by Roediger and McDermott (1995),which revisited an earlier paper published by Deese(1959). This has now become know as the DRM(after Deese, Roediger, and McDermott) paradigm.The procedure typically used in such studiesinvolves presenting a list of 10–15 words for study(REST, AWAKE, DREAM, PILLOW, BED, etc.)that are highly related to a critical nonpresenteditem (SLEEP). The powerful memory illusion hereis that subjects are just as likely to recall (or recog-nize) the critical nonpresented item (SLEEP) as

Semantic Memory 525


items that were actually presented. Moreover, whengiven remember/know judgments (Tulving, 1985),participants often give the critical nonpresenteditem remember judgments that presumably tappeddetailed episodic recollective experience. It is as ifthe strong preexisting semantic memory structure isso powerful that it overwhelms the episodic studyexperience.

It should not be surprising that many of the sameissues that have played out in the semantic memoryresearch have also played out in the false memoryresearch. Indeed, one model in this area is the activa-tion monitoring (AM) framework (e.g., Roedigeret al., 2001a), which suggests that subjects sometimesconfuse the activation that is produced by spreadingactivation that converges on the critical nonpre-sented item (much akin to the Collins and Loftus,1975) with the activation resulting from the studyevent. This framework attempts to keep separatethe episodic and semantic systems but also showshow such systems can interact. In contrast, Arndtand Hirshman (1998) have used the Hintzman(1986) MINERVA framework to accommodate theDRM effect by relying on the similarity of the vec-tors of the individually stored words and the criticalnonpresented items. As noted above, the MINERVAframework does not make a strong distinctionbetween episodic and semantic systems. Moreover,the MINERVAmodel is more a feature-based model,whereas the AM framework a priori would appearmore akin to a prototype model, but no strong claimshave been made along this dimension. A further dis-tinction between the AM framework and theMINERVA approach concerns the relative contribu-tions of backward associative strength (BAS, or theprobability that a list item will elicit the target, orcritical lure, on a free-association task) and forwardassociative strength (FAS, or the probability the crit-ical lure will elicit a list item in such a task).According to AM accounts, the critical variable isexpected to be BAS, as the activation flows from thelist items to the critical lure. However, according toMINERVA, FAS should be more important, as thesimilarity between the probe (i.e., the critical lure)and the stored episodes (i.e., the list items) should bea more powerful determinant of memory perfor-mance. Results from a multiple regression analysisreported by Roediger et al. (2001b) indicated that, inthe DRM paradigm, BAS was the better predictor,thus supporting the AM framework. (We thankRoddy Roediger for pointing this out.)

The question of the nature of the representation(i.e., associative vs semantic) underlying thesepowerful memory illusions has also been studied.For example, Hutchison and Balota (2005) recentlyutilized the summation paradigm developed byBalota and Paul (1996), described earlier, to examinewhether the DRM effect reflects meaning-basedsemantic information or could also be accommo-dated by primarily assuming an associative levelinformation. Hence, in this study, subjects studiedlists of words that were related to one meaning orrelated to two different meanings of a critical non-presented homograph (e.g., the season meaning ofFALL or the accident meaning of FALL). In addi-tion, there were standard DRM lists that onlyincluded words that were related to the same mean-ing of a critical nonpresented word (e.g., such asSLEEP). Consistent with the Balota and Paulresults, the results from both recall and recognitiontests indicated that there was no difference in thepattern of false memory for study lists that con-verged on the same meaning (standard DRM lists)of the critical nonpresented items and lists thatdiverged on different meanings (homograph lists)of the critical nonpresented items. However, whensubjects were required to explicitly make gist-basedresponses and directly access the meaning of the list,that is, is this word related to the studied list, therewas clear (and expected) difference between homo-graph and nonhomograph lists. Hutchison andBalota argued that although rich networks developthrough strategic use of meaning during encodingand retrieval, the activation processes resulting fromthe studied information seem to primarily reflectimplicit associative information and do not demandrich meaning-based analysis.

There is little doubt that what we store in memoryis a reflection of the knowledge base that we alreadyhave in memory, which molds the engram. Hence, asnoted earlier, semantic memories may be episodicmemories that have lost the contextual informationacross time because of repeated exposures. It is un-likely that a 50-year-old remembers the details ofhearing the Rolling Stones’ ‘‘Satisfaction’’ for thefirst time, but it is likely that, soon after that originalexperience, one would indeed have vivid episodicdetails, such as where one was, who one was with,and so on. Although this unitary memory systemapproach clearly has some value (e.g., McKoonet al., 1986), it is also the case that there is somepowerful evidence from cognitive neuroscience thatsupports a stronger distinction.

526 Semantic Memory


2.28.8 Representation andDistinctions: Evidence fromNeuropsychology

Evidence for the distinction of multiple memorysystems has come from studies of patients with loca-lized lesions that produce strong dissociations inbehavior. For example, the classic case of HM (seeScoville and Milner, 1957) indicated that damage tothe hippocampus resulted in impairment of the stor-age of new episodic memories, whereas semanticknowledge appeared to be relatively intact (but seeMacKay et al., 1998). Hence, one might be overlyconcerned about the controversy from the behavioralstudies regarding the distinct nature of semantic andepisodic memory systems. However, there are addi-tional neuropsychological cases that are indeed quiteinformative about the actual nature of semanticrepresentations.

2.28.8.1 Category-Specific Deficits

There have now been numerous cases of individualswho have a specific lesion to the brain and appear tohave localized category-specific deficits. For exam-ple, there have been individuals who have difficultyidentifying items from natural categories (e.g., ani-mals, birds, fruits, etc.) but have a relativelypreserved ability to identify items from artificialcategories (e.g., clothing, tools, furniture). At firstglance, such results would appear to suggest thatcertain categories are represented in distinct neuraltissue that have or have not been disrupted by thelesion. Such a pattern may also be consistent with alocalized representation of meaning instead of a dis-tributed feature-based representation in which allconcepts share vectors of the same set of primitivefeatures.

Unfortunately, however, the interpretation ofimpaired performance on natural categories andintact performance on artificial categories has beencontroversial. For example, such deficits could occurat various stages in the information flow from dis-criminating visually similar items (e.g., Riddoch andHumphreys, 1987) to problems retrieving the appro-priate name of an object (e.g., Hart et al., 1985). Suchaccounts do not rely on the meaning of the categoriesbut suggest that such deficits may reflect correlateddimensions (e.g., difficulty of the visual discrimina-tion) that differ between natural and artificialcategories. In this light, it is particularly important

that there have been cases that have shown the oppo-site pattern. For example, Sacchett and Humphreys(1992) reported an intriguing case that shows disrup-tion of the performance on artificial categories andbody parts but relatively preserved performance onnatural categories. They argued that one possiblereason for this pattern is that this individual had adeficit in representing functional features, which aremore relevant to artifactual representations and bodyparts than natural categories such as fruits and vege-tables. Whatever the ultimate explanation of thesecategory-specific deficits, this work has been infor-mative in providing a better understanding of howmembers within categories may differ on distinctdimensions.

In a similar vein, one hypothesis that hasbeen suggested to explain domain differences incategory-specific deficits is the sensory/functionalhypothesis (Warrington and McCarthy, 1987; Farahand McClelland, 1991; Caramazza and Shelton,1998). According to this proposal, natural categoriessuch as animals depend heavily on perceptual infor-mation (especially on visual discriminations) foridentification and discrimination. Conversely, func-tional information is more important for recognitionof artifacts, such as tools. Thus, damage to regions ofsensory cortex is expected to result in selectiveimpairment of natural kinds, whereas damage toregions in or adjacent to motor cortex would resultin impairment in artifacts. Although compelling, thisview is not endorsed by all researchers. Caramazzaand colleagues, in particular, have argued that thesensory/functional hypothesis fails to account forsome of the patterns of deficits observed and someof the finer-grain distinctions. In particular, it is dif-ficult for this model to account for the selectivesparing or impairment of fruits and vegetables, bodyparts, and musical instruments that have beenreported (see Capitani et al., 2003, for a recentreview). Thus, the question of whether and how thetype of knowledge that is most critical for supportingthe representation of a particular domain is involvedin category-specific deficits remains open.

To address this controversy, Cree and McRae(2003) extended the sensory/functional hypothesis toinclude a broader range of types of knowledge. Theydeveloped a brain region taxonomy that included ninedifferent forms of knowledge, including sensory/perceptual in all modalities (vision, taste, audition,etc.), functional, and encyclopedic. Encyclopedic fea-tures included information about items such as LIVESIN AFRICA for ELEPHANT – in other words,

Semantic Memory 527


information that likely was learned and not experi-enced directly. Cree and McRae then developed anine-dimensional representation for the 541 conceptsfor which they had norming data and estimated thesalience of each type of knowledge for each object andeach category. In a series of cluster analyses, Cree andMcRae found that the knowledge types nicely pre-dicted the tripartite distinction between living things,artifacts, and fruits and vegetables reported in severalneuropsychological case studies. In addition, Cree andMcRae examined several distributional statistics,including the number of distinguishing and distinctivefeatures and similarity to obtain a measure of confu-sability (i.e., the extent to which a given concept mightbe confused with another concept from the same cate-gory). The categories they examined did appear to bedifferentially sensitive to these measures, and theimplemented model reflected patterns of impairmentobserved in patients. They concluded that knowledgetype does underlie the organization of conceptualrepresentations and that selective impairment in aparticular brain region involved in maintaining suchknowledge can result in the observed patterns ofimpairment in patients with category-specific deficits.Although many questions remain, it is clear that evi-dence from individuals with category-specific deficitshas provided considerable insight into both the natureof category representation and the underlying neuralrepresentations.

2.28.8.2 Semantic Dementia

The most common form of dementing illness isdementia of the Alzheimer type (DAT). However,there is also a relatively rare and distinct dementia,referred to as semantic dementia (SD), which over-laps with DAT in features such as insidious onset andgradual deterioration of comprehension and word-finding ability. SD is a variant of frontal temporaldementia and typically involves one or both of theanterior portions of the temporal lobes. The consen-sus criteria for SD (Hodges et al., 1992) includeimpairment in semantic memory causing anomia,deficits in both spoken and written word comprehen-sion, a reading pattern consistent with surfacedyslexia (i.e., impairment in reading exceptionwords such as PINT but preserved reading of regularwords and nonwords that follow standard spelling tosound rules, such as NUST), impoverished knowl-edge about objects and/or people with relativesparing of phonological and syntactic componentsof speech output, and perceptual and nonverbal

problem solving skills. These individuals are oftenquite fluent, but their speech is relatively limited inconveying meaning. They are particularly poor atpicture naming and understanding the relationsamong objects. For example, the Pyramids andPalm Trees test developed by Howard andPatterson (1992) involves selecting which of twoitems (e.g., a palm tree or a fir tree) is most similarto a third item (e.g., a pyramid). Individuals with SDare particularly poor at this task and so would appearto have a breakdown in the representations of theknowledge structures.

An interesting dissociation has been madebetween SD individuals and DAT individuals. Inparticular, Simons et al. (2002) recently found a dou-ble dissociation, wherein individuals with SDproduced poorer picture naming than individualswith DAT; however, individuals with SD producedbetter performance than individuals with DAT on alater episodic recognition test of these very samepictures (also see Gold et al., 2005). Clearly, theselective impairment across these two groups ofparticipants is consistent with distinct types of infor-mation driving these tasks. Of course, one must becautious about the implications even from this study,because it is unlikely that either task is a process-puremeasure of episodic and semantic memory (seeJacoby, 1991), but clearly these results are veryintriguing.

Recently, Rogers et al. (2004) proposed a model ofsemantic memory that maintains strong connectionsto modality-specific systems in terms of both inputsand outputs and has been particularly useful inaccommodating the deficits observed in SD. Thismodel has some interesting parallels to Barsalou’s(1999) proposal, in that it assumes that semanticmemory is grounded in perception and action net-works. In addition, like the model proposed byMcRae et al. (1997), Rogers et al. suggest that thesystem is sensitive to statistical regularities, and theseregularities are what underlie the development ofsemantics. The particular contribution of Rogerset al.’s model, however, is that although semanticrepresentations are grounded in perception–actionmodality-specific systems, the statistical learningmechanism allows the emergence of abstract seman-tic representations. Importantly, inputs to semanticsare mediated by perceptual representations that aremodality specific, and as a result, the content ofsemantic memory relies on the same neural tissuethat supports encoding. However, different fromBarsalou and colleagues’ account, Rogers et al. do

528 Semantic Memory


suggest that there is a domain-general, abstractedrepresentation that emerges from cross-modal map-pings. Thus, although the system relies on perceptualinputs, the abstract representations can capture cross-modality similarities and structures to give rise tosemantic memory.

Rogers et al. (2004) implemented a simple versionof their model using a parallel distributed-processingapproach in which visual features provided the percep-tual input and are allowed to interact in training withverbal descriptors through a mediating semantic level.Importantly, the semantic representations emergethrough the course of training as the network learnsthe mappings between units at the visual and verballevels. The units the model was trained on consisted ofverbal and visual features generated in separate norm-ing sessions. Once training was complete, severalsimulations were reported in which the model wasprogressively damaged in a way that was thought tomimic varying levels of impairment observed in indi-viduals with SD. Overall, the model nicely capturedthe patterns of performance of the patients.Specifically, one pattern often observed in SD is atendency to overregularize conceptual knowledge.For example, individuals might refer to all exemplarsof a category using the superordinate label or a singlelabel that is high in frequency (e.g., calling a DOG anANIMAL or a ZEBRA a HORSE). This is possibly aresult of the progressive failure in retrieving idiosyn-cratic information that serves to distinguish exemplars,such that only the central tendency (e.g., a prototype ormost typical exemplar) remains accessible. Thus, lesscommon items might take on the attributes of higher-frequency exemplars. The model displayed similarpatterns of generalization as the SD individuals, afinding explained in terms of changes in attractordynamics that resulted in the relative sparing of fea-tures and attributes shared by many exemplars but aloss of more distinctive features. This model providesan interesting account of semantic memory and thedeficits observed in individuals with SD, one in whichboth perceptually based information and abstractedrepresentations interact to give rise to knowledge ofthe world.

2.28.9 Neuroimaging

Investigations into the nature of semantic memoryhave benefited from recent advances in technologythat allow investigators to examine online processingof information in the human brain. For example,

positron emission technology (PET) and functionalmagnetic resonance imaging (fMRI) allow one to mea-sure correlates of neural activity in vivo as individualsare engaged in semantic tasks (see Logothetis andWandell, 2004). Although a full review of the substan-tial contributions of neuroimaging data to the questionspertaining to semantics is beyond the scope of thischapter (See Chapter 3.07 for a review), we brieflyexamine some of the major findings that have helpedconstrain recent theorizing about the nature and locusof semantic representations. Two major brain regionshave been identified through neuroimaging studies: leftprefrontal cortex (LPC) and areas within the temporallobes, particularly in the left hemisphere.

The first study to report neuroimaging data rele-vant to semantic memory was conducted by Petersenet al. (1988), who used PET techniques to localizeactivation patterns specific to semantic tasks. Subjectswere asked to generate action verbs upon presenta-tion of a concrete object noun, and activity duringthis task was compared with the activity occurringduring silent reading of the words. Petersen et al.reported significant patterns of activity in LPC, afinding that has since been replicated and extendedto other types of attributes. Martin et al. (1995)extended this work to show that the specific attributeto be retrieved yielded different patterns of activa-tion. Specifically, the locus of activation involved inattribute retrieval tends to be in close proximity tothe neural regions that are involved in perception ofthe specific attributes. Thus, retrieval of visual infor-mation, such as color, tends to activate regionsadjacent to the regions involved in color perception,whereas retrieval of functional information results inactivation of areas adjacent to motor cortex. Thesefindings mesh nicely with the perceptual/motornotions of representation in semantic memoryreviewed above (e.g., Barsalou, 1999; Rogers et al.,2004). In addition, Roskies et al. (2001) reported thatnot only were regions in lateral inferior prefrontalcortex (LIPC) preferentially active during tasks thatrequired semantic processing, but specific regionswere also sensitive to task difficulty. Thus, it appearsthat frontal regions are involved both in the activeretrieval from semantic memory and in processingspecific semantic information.

Many researchers have suggested, however, thatalthough frontal regions are involved in semantic re-trieval, the storage of semantic information is primarilyin the temporal regions (see Hodges et al., 1992).Indeed, another area that has been implicated insemantic processing is in the ventral region of the

Semantic Memory 529


temporal lobes, centered on the fusiform gyrus, andespecially in the left hemisphere. This area showssignificant activation during word reading and objectnaming tasks, indicating it is not sensitive to the stim-ulus form but to the semantic content therein (seeMartin, 2005, for a review). Furthermore, within thisarea, different subregions become more or less acti-vated when subjects view faces, houses, and chairs (e.g.,Chao et al., 1999), suggesting that different domainsrely on different regions of neural tissue. This, ofcourse, could be viewed as consistent with the cate-gory-specific deficits reviewed above. However, asnoted by Martin and Chao (2001), although peak acti-vation levels in response to objects from differentdomains reflect a certain degree of localization, thepredominant finding is a pattern of broadly distributedactivation throughout the ventral temporal and occip-ital regions, which is consistent with the idea thatrepresentations are distributed over large corticalregions.

Recently, Wheatley et al. (2005) reported datafrom a semantic priming study using fMRI that alsoconverges on the notion of perceptual motor repre-sentations of meaning. Subjects silently read related,unrelated, or identical word pairs at a 250-ms SOAwhile being scanned. The related pairs consisted ofcategory members that were not strongly associa-tively related (e.g., DOG-GOAT, but see thediscussion above regarding the difficulty of selectingsuch items). Given the relatively fast SOA and that noovert response was required, Wheatley et al. arguedthat any evidence for priming should be a reflection ofautomatic processes. Consistent with other evidencethat indicates there are reliable neural correlates ofbehavioral priming that were evidenced by reducedhemodynamic activity (Wiggs and Martin, 1998;Mummery et al., 1999; Rissman et al., 2003;Maccotta and Buckner, 2004), Wheatley et al. founddecreased activity for identity pairs and a slightlysmaller, but still significant, decrease for relatedpairs relative to the unrelated pairs condition.Importantly, Wheatley et al. were able to comparepatterns of activation as a function of domain.Consistent with proposals by Barsalou (1999), theyfound that objects from animate objects yielded moreactivity in regions adjacent to sensory cortex, whereasmanipulable artifacts resulted in greater activity inregions adjacent to motor cortex. These findingswere taken as evidence that conceptual informationabout objects is stored, at least in part, in neuralregions that are involved in perception and action.

Although the Wheatley et al. (2005) study used atask that was likely to minimize strategic processing,one question that remains to be addressed is whetherthe automatic and strategic processes involved insemantic priming tasks (see earlier discussion) canalso be dissociated in neural tissue. In a recentstudy, Gold et al. (2006) reported that several of thebrain regions previously implicated in processingduring semantic tasks are differentially sensitive tothe automatic and strategic processes involved inlexical decision tasks. In three experiments, Goldet al. manipulated prime target relatedness, SOA,and whether primes and targets were orthographi-cally or semantically related. Long and short SOAswere intermixed in scanning runs to assess the rela-tive contributions of strategic and automaticprocesses (see Neely, 1991). A comparison of ortho-graphic and semantic priming conditions wasincluded to determine whether any areas were par-ticularly sensitive to the two sources of priming orwhether priming effects are more general mecha-nisms. The results clearly indicated that differentregions responded selectively to different conditions.Specifically, midfusiform gyrus was more sensitive toautomatic than strategic priming, but only for seman-tically related primes, as this region did not showreduced activity for orthographic primes. Fourregions were more sensitive to strategic than auto-matic priming, two in left anterior prefrontal cortexand bilateral anterior cingulate. Even more intrigu-ing, the two regions in LIPC were further dissociated:The anterior region showed strategic semantic facil-itation, as evidenced by decreased activity, relative toa neutral baseline, whereas the posterior regionshowed strategic semantic inhibition, or increasedactivity, relative to the neutral baseline. In addition,the medial temporal gyrus showed decreased activa-tion concurrently with the anterior LIPC, supportingprevious claims that these regions show greater acti-vation in tasks that are more demanding of strategicprocesses but reduced activation when the strategicprocesses are less demanding (Wagner et al., 2000;Gold et al., 2005). In sum, it appears that the behav-ioral dissociations between automatic and strategicprocesses in priming tasks are also found in the neu-roimaging data. The complexity of the patterns ofactivation involved in semantic tasks appears to indi-cate that the retrieval and storage of semanticinformation is indeed a distributed phenomenonthat requires the coordination of a wide array ofneural tissue.

530 Semantic Memory


2.28.10 Development andBilingualism

Although we have attempted to provide a review ofthe major issues addressed in semantic memoryresearch, there are clearly other important areas thatwe have not considered in detail because of lengthlimitations. For example, there is a very rich area ofdevelopmental research addressing the acquisition ofmeaning in children (see Bloom, 2000, for a compre-hensive review), along with work that attempts tocapture the nature of semantic memory in olderadulthood (see, e.g., Balota and Duchek, 1989). Ofcourse, we touched upon these issues earlier whendiscussing how the small world networks of Steyversand Tenenbaum (2005) develop over time, along withthe work by Rosch (1975) on the development ofcategorization. Given that meaning is extracted frominteractions with the environment, the developmentalliterature is particularly important to understand howadditional years of experience mold the semanticsystem, especially in very early life. There are manyinteresting connections of this work to topics coveredearlier in this chapter. For example, regarding theinfluence of preexisting structures on false memory,it is noteworthy that young children (5-year-olds) aremore likely to produce phonological than semanticfalse memories, whereas older children (around 11years and older) are more likely to produce the oppo-site pattern (see Dewhurst and Robinson, 2004).Possibly, this is a natural consequence of the devel-opment of a rich semantic network in early childhoodthat lags behind a more restricted phonologicalsystem.

Another very active area of research involves thenature of semantic representations in bilinguals (seeFrancis, 1999, 2005, for excellent reviews). For exam-ple, researchers have attempted to determine whetherthere is a common semantic substrate that is amodal,with each language having specific lexical level repre-sentations (e.g., phonology, orthography, syntax, etc.)that map onto this system. This contrasts with theview that each language engages distinct semanticlevel representations. Although there is still somecontroversy, the experimental results seem more con-sistent with the assumption that the semantic level isshared across languages, at least for skilled bilinguals.Evidence in support of this claim comes from adiverse range of tasks. For example, in a mixed lan-guage list, memory for the language of input isgenerally worse than memory for the concepts

(e.g., Dalrymple-Alford and Aamiry, 1969). In addi-tion, one finds robust semantic priming effects bytranslation equivalents (words in different languageswith the same meaning, e.g., DOG in English andHUND in German), which is consistent with at leasta partially shared semantic representation (e.g., deGroot and Nas, 1991; Gollan et al., 1997).

2.28.11 Closing Comments

The nature of how humans develop, represent, andefficiently retrieve information from their vast repo-sitory of knowledge has for centuries perplexedinvestigators of the mind. Although there is clearlyconsiderable work to be done, recent advances inanalyses of large-scale databases, new theoreticalperspectives from embodied cognition and smallworld networks, and new technological develop-ments allowing researchers to measure, in vivo,brain activity, are making considerable progresstoward understanding this fundamental aspect ofcognition.

References

Albert R and Barabasi AL (2000) Topology of evolving networks:Local events and universality. Phys. Rev. Lett. 85:5234–5237.

Anderson JR (2000) Learning and Memory: An IntegratedApproach, 2nd edn. Pittsburgh: Carnegie Mellon.

Anderson JR and Bower GH (1973) Human AssociativeMemory. Washington DC: Hemisphere Press.

Anderson MC, Bjork RA, and Bjork EL (1994) Remembering cancause forgetting: Retrieval dynamics in long-term memory.J. Exp. Psychol. Learn. Mem. Cogn. 20: 1063–1087.

Arndt J and Hirshman E (1998) True and false recognition inMINERVA2: Explanations of a global matching perspective.J. Mem. Lang. 39: 371–391.

Atkinson RC and Juola JF (1974) Search and decisionprocesses in recognition memory. In: Krantz DH,Atkinson RC, et al. (eds.) Contemporary Developments inMathematical Psychology: I Learning, Memory, andThinking, pp. 243–293. San Francisco: Freeman.

Baddeley A (2000) Short-term and working memory.In: Tulving E and Craik FIM (eds.) The Oxford Handbook ofMemory, pp. 77–92. Oxford: Oxford University Press.

Balota DA and Chumbley JI (1984) Are lexical decisions a goodmeasure of lexical access? The role of word frequency in theneglected decision stage. J Exp. Psychol. Hum. Percept.Perform. 10: 340–357.

Balota DA and Duchek JM (1989) Spreading activation inepisodic memory: Further evidence for age-independence.Q. J. Exp. Psychol. 41A: 849–876.

Balota DA and Lorch RF (1986) Depth of automatic spreadingactivation: Mediated priming effects in pronunciation but notin lexical decision. J. Exp. Psychol. Learn. Mem. Cogn. 12:336–345.

Semantic Memory 531


Balota DA and Paul ST (1996) Summation of activation:Evidence from multiple primes that converge and divergewithin semantic memory. J. Exp. Psychol. Learn. Mem.Cogn. 22: 827–845.

Balota DA, Cortese MJ, Sergent-Marshall S, Spieler DH, andYap MJ (2004) Visual word recognition of single syllablewords. J. Exp. Psychol. Gen. 133: 336–345.

Barsalou LW (1985) Ideals, central tendency, and frequency ofinstantiation as determinants of graded structure incategories. J. Exp. Psychol. Learn. Mem. Cogn. 11: 629–654.

Barsalou LW (1999) Perceptual symbol systems. Behav. BrainSci. 22: 577–609.

Barsalou LW (2003) Abstraction in perceptual symbol systems.Philos. Trans. R. Soc. Lond. Biol. Sci. 358: 1177–1187.

Barsalou LW, Simmons WK, Barbey A, and Wilson CD (2003)Grounding conceptual knowledge in modality-specificsystems. Trends Cogn. Sci. 7: 84–91.

Bartlett FC (1932) Remembering. Cambridge: CambridgeUniversity Press.

Battig WF and Montague WE (1969) Category norms for verbalitems in 56 categories: A replication and extension of theConnecticut category norms. J. Exp. Psychol. Monogr. 80:1–46.

Bloom P (2000) How Children Learn the Meanings of Words.Cambridge MA: MIT Press.

Bousfield WA (1953) The occurrence of clustering in the recall ofrandomly arranged associates. J. Gen. Psychol. 49: 229–240.

Bruce D and Fagan RL (1970) More on the recognition and freerecall of organized lists. J. Exp. Psychol. 85: 153–154.

Bruner JS, Goodnow JJ, and Austin GA (1956) A Study ofThinking. New York: Wiley.

Burgess C and Lund K (1997) Modeling parsing constraints withhigh-dimensional context space. Lang. Cogn. Process. 12:177–210.

Capitani E, Laiacona M, Mahon B, and Caramazza A (2003)What are the facts of semantic category-specific deficits? Acritical review of the clinical evidence. Cogn. Neuropsychol.20: 213–261.

Caramazza A and Shelton JR (1998) Domain-specificknowledge systems in the brain: The animate-inanimatedistinction. J. Cogn. Neurosci. 10: 1–34.

Carey S (1978) The child as a word learner. In: Halle M, Bresnan J,and Miller GA (eds.) Linguistic Theory and PsychologicalReality, pp. 264–293. Cambridge MA: MIT Press.

Chao LL, Haxby JV, and Martin A (1999) Attribute-based neuralsubstrates in temporal cortex for perceiving and knowingabout objects. Nat. Neurosci. 2: 913–919.

Chwilla DJ and Kolk HHJ (2002) Three-step priming in lexicaldecision. Mem. Cognit. 30: 217–225.

Cofer CN, Bruce DR, and Reicher GM (1966) Clustering in freerecall as a function of certain methodological variations. J.Exp. Psychol. 71: 858–866.

Collins AM and Loftus EF (1975) A spreading-activation theoryof semantic processing. Psychol. Rev. 82: 407–428.

Collins A and Quillian MR (1969) Retrieval time from semanticmemory. J. Verb. Learn. Verb. Behav. 8: 240–247.

Cree GS and McRae K (2003) Analyzing the factors underlyingthe structure and computation of the meaning of chipmunk,cherry, chisel, cheese and cello (and many other suchconcrete nouns). J. Exp. Psychol. Gen. 132: 163–201.

Cree GS, McRae K, and McNorgan C (1999) An attractor modelof lexical conceptual processing: Simulating semanticpriming. Cogn. Sci. 23: 371–414.

Crowder RG (1976) Principles of Learning and Memory.Hillsdale, NJ: Erlbaum.

Dalrymple-Alford EC and Aamiry A (1969) Word associations ofbilinguals. Psychon. Sci. 21: 319–320.

Deese J (1959)On theprediction of occurrence of particular verbalintrusions in immediate free recall. J. Exp. Psychol. 58: 17–22.

de Groot AMB (1983) The range of automatic spreadingactivation in word priming. J. Verb. Learn. Verb. Behav. 22:417–436.

de Groot AMB and Nas GLJ (1991) Lexical representation ofcognates and non-cognates in compound bilinguals. J.Mem. Lang. 30: 90–123.

de Mornay Davies P (1998) Automatic semantic priming: Thecontribution of lexical- and semantic-level processes. Eur. J.Cogn. Psychol. 10: 398–412.

Dewhurst SA and Robinson CA (2004) False memories inchildren: Evidence for a shift from phonological to semanticassociations. Psychol. Sci. 15: 782–786.

Durgunoglu AY and Neely JH (1987) On obtaining episodicpriming in a lexical decision task following paired-associatelearning. J. Exp. Psychol. Learn. Mem. Cogn. 13: 206–222.

Ebbinghaus H (1885) On Memory. Ruger HA and Bussenius CE(trans.) New York: Teacher’ College, 1913.

Faust ME, Balota DA, and Spieler DH (2001) Building episodicconnections: Changes in episodic priming with age anddementia. Neuropsychology 15: 626–637.

Farah M and McClelland JL (1991) A computational model ofsemantic memory impairment: Modality-specificity andemergent category specificity. J. Exp. Psychol. Gen. 122:339–357.

Fischler I (1977) Semantic facilitation without association in alexical decision task. Mem. Cognit. 5: 335–339.

Fodor J (1983) The Modularity of Mind. Cambridge MA: MITPress.

Francis WS (1999) Cognitive integration of language andmemory in bilinguals: Semantic representation. Psychol.Bull. 125: 193–222.

Francis WS (2005) Bilingual semantic and conceptualrepresentation. In: Kroll JF and de Groot AMB (eds.)Handbook of Bilingualism: Psycholinguistic Approaches,pp. 251–267. Oxford: Oxford University Press.

Gelman SA and Markman EM (1986) Categories and inductionin young children. Cognition 23: 183–208.

Gibson JJ (1979) The Ecological Approach to Visual Perception.Hillsdale, NJ: Erlbaum.

Glenberg AM and Kaschak MP (2002) Grounding language inaction. Psychon. Bull. Rev. 9: 558–565.

Glenberg AM and Robertson DA (2000) Symbol grounding andmeaning: A comparison of high-dimensional and embodiedtheories of meaning. J. Mem. Lang. 43: 379–401.

Gold BT, Balota DA, Cortese MJ, et al. (2005) Differingneuropsychological and neuroanatomical correlates ofabnormal reading in early-stage semantic dementia anddementia of the Alzheimer type. Neuropsychologia 43:833–846.

Gold BT, Balota DA, Jones SJ, Powell DK, Smith CD, andAndersen AH (2006) Dissociation of automatic and strategiclexical-semantics: fMRI evidence for differing roles ofmultiple frontotemporal regions. J. Neurosci. 26: 6523–6532.

Gollan TH, Forster KI, and Frost R (1997) Translation primingwith different scripts: Masked priming with cognates andnoncognates in Hebrew-English bilinguals. J. Exp. Psychol.Learn. Mem. Cogn. 23: 1122–1139.

Hart J, Berndt RS, and Caramazza A (1985) Category-specificnaming deficit following cerebral infarction. Nature 316:439–440.

Herrnstein RJ, Loveland DH, and Cable C (1976) Naturalconcepts in pigeons. J. Exp. Psychol. Anim. Behav. Process.2: 285–302.

Hines D, Czerwinski M, Sawyer PK, and Dwyer M (1986)Automatic semantic priming: Effect of category exemplarlevel and word association level. J. Exp. Psychol. Hum.Percept. Perform. 12: 370–379.

Hintzman DL (1986) ‘‘Schema abstraction’’ in a multiple tracememory model. Psychol. Rev. 93: 411–428.

532 Semantic Memory


Hintzman DL (1988) Judgements of frequency and recognitionmemory in a multiple-trace memory model. Psychol. Rev. 95:528–551.

Hodges JR, Patterson K, Oxbury S, and Funnell E (1992)Semantic dementia: Progressive fluent aphasia withtemporal lobe atrophy. Brain 115: 1793–1806.

Hoffman AB and Murphy GL (2006) Category dimensionalityand feature knowledge: When more features are learned aseasily as fewer. J. Exp. Psychol. Learn. Mem. Cogn. 32:301–315.

Hoffman RR, Bringmann W, Bamberg M, and Klein R (1987)Some historical observations on Ebbinghaus. In: Gorfein Dand Hoffman R (eds.)Memory and Learning: The EbbinghausCentennial Conference, pp. 57–76. Hillsdale, NJ: Erlbaum.

Howard D and Patterson K (1992) The Pyramids and Palm TreesTest. Bury St. Edmunds: Thames Valley Test Company.

Hutchison KA (2003) Is semantic priming due to associationstrength or feature overlap? A micro-analytic review.Psychon. Bull. Rev. 10: 785–813.

Hutchison KA and Balota DA (2005) Decoupling semantic andassociative information in false memories: Explorations withsemantically ambiguous and unambiguous critical lures. J.Mem. Lang. 52: 1–28.

Jacoby LL (1991) A process dissociation framework: Separatingautomatic from intentional uses of memory. J. Mem. Lang.30: 513–541.

Jeong H, Kahng B, Lee S, Kwak CY, Barabasi AL, and FurdynaJK (2000) Monte Carlo simulation of sinusoidally modulatedsuperlattice growth. Phys. Rev. E Stat. Nonlin. Soft MatterPhys. 65: 031602–031605.

LaBerge D and Samuels SJ (1974) Toward a theory of automaticinformation processing in reading.Cogn. Psychol. 6: 293–323.

Landauer TK and Dumais ST (1997) A solution to Plato’sproblem: The latent semantic analysis theory of acquisition,induction, and representation of knowledge. Psychol. Rev.104: 211–240.

Leach E (1964) Anthropological aspects of language: Animalcategories and verbal abuse. In: Lenneberg EH (ed.) NewDirections in the Study of Language. Cambridge, MA: MITPress.

Livesay K and Burgess C (1998) Mediated priming in high-dimensional semantic space: No effect of direct semanticrelationships or co-occurrence. Brain Cogn. 37: 102–105.

Logothetis NK and Wandell BA (2004) Interpreting the BOLDsignal. Annu. Rev. Physiol. 66: 735–769.

Lucas M (2000) Semantic priming without association: A meta-analytic review. Psychon. Bull. Rev. 7: 618–630.

Lupker SJ (1984) Semantic priming without association: Asecond look. J. Verb. Learn. Verb. Behav. 23: 709–733.

Maccotta L and Buckner RL (2004) Evidence for neural effectsof repetition that directly correlate with behavioral priming. J.Cogn. Neurosci. 16: 1625–1632.

MacKay DG, Burke DM, and Stewart R (1998) HM’s languageproduction deficits: Implications for relationships betweenmemory, semantic binding, and the hippocampal system. J.Mem. Lang. 38: 28–69.

Martin A (2005) Functional neuroimaging of semantic memory.In: Cabeza R and Kingstone A (eds.) Handbook of FunctionalNeuroimaging of Cognition, pp. 153–186. Cambridge, MA:MIT Press.

Martin A and Chao LL (2001) Semantic memory and the brain:Structure and processes. Curr. Opin. Neurobiol. Cogn.Neurosci. 11: 194–201.

Martin A, Haxby JV, Lalonde FM, Wiggs CL, and Ungerleider LG(1995) Discrete cortical regions associated with knowledgeof color and knowledge of action. Science 270: 102–105.

McCloskey M and Glucksberg S (1979) Decision processes inverifying category membership statements: Implications formodels of semantic memory. Cogn. Psychol. 11: 1–37.

McKoon G and Ratcliff R (1979) Priming in episodic andsemantic memory. J. Verb. Learn. Verb. Behav. 18: 463–480.

McKoon G and Ratcliff R (1992) Spreading activation versuscompound cue accounts of priming: Mediated primingrevisited. J. Exp. Psychol. Learn. Mem. Cogn. 18:1155–1172.

McKoon G, Ratcliff R, and Dell GS (1986) A critical evaluation ofthe semantic episodic distinction. J. Exp. Psychol. Learn.Memo. Cogn. 12: 295–306.

McNamara TP and Altarriba J (1988) Depth of spreadingactivation revisited: Semantic mediated priming occurs inlexical decisions. J. Mem. Lang. 27: 545–559.

McRae K (2004) Semantic memory: Some insights from feature-based connectionist attractor networks. In: Ross BH (ed.)Psychology of Learning and Motivation, vol. 45, pp. 41–86.San Diego: Elsevier.

McRae K, De Sa VR, and Seidenberg MS (1997) On the natureand scope of featural representations of word meaning. J.Exp. Psychol. Gen. 126: 99–130.

McRae K, Cree GS, Westmacott R, and de Sa VR (1999) Furtherevidence for feature correlations in semantic memory. Can.J. Exp. Psychol. (Special issue on models of wordrecognition) 53: 360–373.

McRae K, Cree GS, Seidenberg MS, and McNorgan C (2005)Semantic feature production norms for a large set of livingand nonliving things. Behav. Res. Methods Instr. Comput.37: 547–559.

Medin DL (1989) Concepts and conceptual structure. Am.Psychol. 44: 1469–1481.

Meyer DE and Schvaneveldt RW (1971) Facilitation inrecognizing words: Evidence of a dependence upon retrievaloperations. J. Exp. Psychol. 90: 227–234.

Milgram S (1967) The small world problem. Psychol. Today 2:60–67.

Miller GA (1990) WordNet: An on-line lexical database. Int. J.Lexiogr. 3: 235–312.

Mummery CJ, Shallice T, and Price CJ (1999) Dual-processmodel in semantic priming: A functional neuroimagingperspective. Neuroimage 9: 516–525.

Murphy G (2002) The Big Book of Concepts. Cambridge, MA:MIT Press.

Neely JH (1977) Semantic priming and retrieval from lexicalmemory: Roles of inhibitionless spreading activation andlimited-capacity attention. J. Exp. Psychol. Gen. 106:226–254.

Neely JH (1991) Semantic priming effects in visual wordrecognition: A selective review of current findings andtheories. In: Besner D and Humphreys G (eds.) BasicProcesses in Reading: Visual Word Recognition,pp. 236–264. Hillsdale, NJ: Erlbaum.

Neely JH and Durgunoglu AY (1985) Dissociative episodic andsemantic priming effects in episodic recognition and lexicaldecision tasks. J. Mem. Lang. 24: 466–489.

Nelson DL, McEvoy CL, and Schreiber TA (1998) The Universityof South Florida word association, rhyme, and wordfragment norms. http://www.usf.edu/FreeAssociation/.

Osgood CE, Suci GJ, and Tanenbaum PH (1957) TheMeasurement of Meaning. Urbana: University of IllinoisPress.

Pecher D and Raaijmakers JGW (1999) Automatic primingeffects for new associations in lexical decision andperceptual identification. Q. J. Exp. Psychol. 52A: 593–614.

Pecher D, Zeelenberg R, and Barsalou LW (2003) Verifyingproperties from different modalities for concepts producesswitching costs. Psychol. Sci. 14: 119–124.

Pecher D, Zeelenberg R, and Barsalou LW (2004) Sensorimotorsimulations underlie conceptual representations:Modality-specific effects of prior activation. Psychon. Bull.Rev. 11: 164–167.

Semantic Memory 533


Petersen SE, Fox PT, Posner M, Mintun M, and Raichle M(1988) Positron emission topographic studies of the corticalanatomy of single word processing. Nature 331: 585–589.

Pexman PM, Lupker SJ, and Hino Y (2002) The impact offeedback semantics in visual word recognition: Number offeatures effects in lexical decision and naming tasks.Psychon. Bull. Rev. 9: 542–549.

Pexman PM, Holyk GG, and Monfils M-H (2003) Number offeatures effects and semantic processing. Mem. Cognit. 31:842–855.

Pinker S (1994) The Language Instinct. New York: HarperCollins.

Posner MI and Keele SW (1968) On the genesis of abstractideas. J. Exp. Psychol. 77: 353–363.

Posner MI and Snyder CRR (1975) Attention and cognitivecontrol. In: Solso R (ed.) Information Processing andCognition: The Loyola Symposium, pp. 55–85. Hillsdale, NJ:Erlbaum.

Quillian M (1968) Semantic memory. In: Minsky M (ed.)Semantic Information Processing, pp. 227–270. Cambridge,MA: MIT Press.

Quine WVO (1960) Word and Object. Cambridge, MA: MITPress.

Riddoch MJ and Humphreys GW (1987) Visual objectprocessing in optic aphasia: A case of semantic accessagnosia. Cogn. Neuropsychol. 4: 131–185.

Rissman J, Eliassen JC, and Blumstein SE (2003) An event-related fMRI investigation of implicit semantic priming. J.Cogn. Neurosci. 15: 1160–1175.

Roediger HL III and McDermott KB (1995) Creating falsememories: Remembering words not presented in lists. J.Exp. Psychol. Learn. Mem. Cogn. 21: 803–814.

Roediger HL III, Balota DA, and Watson JM (2001a) Spreadingactivation and arousal of false memories. In: Roediger HLand Nairne JS (eds) The Nature of Remembering: Essays inHonor of Robert G. Crowder, pp. 95–115. Washington DC:American Psychological Association.

Roediger HL III, Watson JM, McDermott KB, and Gallo DA(2001b) Factors that determine false recall: A multipleregression analysis. Psychon. Bull. Rev. 8: 385–407.

Rogers TT, Lambon RMA, Garrard P, et al. (2004) Structure anddeterioration of semantic memory: A neuropsychologicaland computational investigation. Psychol. Rev. 111:205–235.

Roget PM (1911) Roget’s Thesaurus of English Words andPhrases. Available from Project Gutenberg IllinoisBenedictine College, Lisle, IL.

Rosch E (1973) On the internal structure of perceptual andsemantic categories. In: Moore TE (ed.) CognitiveDevelopment and the Acquisition of Language, pp. 111–144.New York: Academic Press.

Rosch E (1975) Cognitive representations of semanticcategories. J. Exp. Psychol. Gen. 104: 192–233.

Rosch E, Mervis CB, Gray WD, Johnson DM, and Boyes-BraemP (1976) Basic objects in natural categories. Cogn. Psychol.8: 382–440.

Roskies AL, Fiez JA, Balota DA, and Petersen SE (2001)Task-dependent modulation of regions in left inferior

frontal cortex during semantic processing. J. Cogn.Neurosci. 13: 1–16.

Sachett C and Humphreys GW (1992) Calling a squirrel asquirrel but a canoe a wigwam: A category-specific deficitfor artefactual objects and body parts. Cogn. Neuropsychol.9: 73–86.

Sayette MA, Hufford MR, and Thorson GM (1996) Developing abrief measure of semantic priming. J. Clin. Exp.Neuropsychol. 8: 678–684.

Scoville WB and Milner B (1957) Loss of recent memory afterbilateral hippocampal lesions. J. Neurol. Neurosurg.Neuropsychiatry 20: 11–21.

Simons JS, Graham KS, and Hodges JR (2002) Perceptual andsemantic contributions to episodic memory: Evidence fromsemantic dementia and Alzheimer’s disease. J. Mem. Lang.47: 197–213.

Smith EE, Shoben EJ, and Rips LJ (1974) Structure and processin semantic memory: A featural model for semanticdecisions. Psychol. Rev. 81: 214–241.

Solomon KO and Barsalou LW (2004) Perceptual simulation inproperty verification. Mem. Cognit. 32: 244–259.

Sperling G (1960) The information available in brief visualpresentation. Psychol. Monogr. 74(11) (Whole No. 498).

Spieler DH and Balota DA (1996) Characteristics of associativelearning in younger and older adults: Evidence from anepisodic priming paradigm. Psychol. Aging 11: 607–620.

Steyvers M and Tenenbaum JB (2005) The large-scale structureof semantic networks: Statistical analyses and a model ofsemantic growth. Cogn. Sci. 29: 41–78.

Thompson-Schill SL, Kurtz KJ, and Gabrieli JDE (1998) Effectsof semantic and associative relatedness on automaticpriming. J. Mem. Lang. 38: 440–458.

Tulving E (1972) Episodic and semantic memory. In: Tulving Eand Donaldson W (eds.) Organization of Memory,pp. 381–403. New York: Academic Press.

Tulving E (1985) Memory and consciousness. Can. Psychol. 26:1–12.

Tulving E and Pearlstone Z (1966) Availability versusaccessibility of information in memory for words. J. Verb.Learn. Verb. Behav. 5: 381–391.

Wagner AD, Koutstaal W, Maril A, Schacter DL, and Buckner RL(2000) Task-specific repetition priming in left inferiorprefrontal cortex. Cereb. Cortex 10: 1176–1184.

Warrington EK and McCarthy R (1987) Categories ofknowledge: Further fractionations and an attemptedintegration. Brain 110: 1273–1296.

Watts DJ and Strogatz SH (1998) Collective dynamics of ‘‘small-world’’ networks. Nature 393: 440–442.

Wheatley T, Weisberg J, Beauchamp MS, and Martin A (2005)Automatic priming of semantically related words reducesactivity in the fusiform gyrus. J. Cogn. Neurosci. 17:1871–1885.

Wickens DD (1973) Some characteristics of word encoding.Mem. Cogn. 1: 485–490.

Wiggs CL and Martin A (1998) Properties and mechanisms ofperceptual priming. Curr. Opin. Neurobiol. 8: 227–233.

Wilson M (2002) Six views of embodied cognition. Psychon.Bull. Rev. 9: 625–636.

534 Semantic Memory

Documents

2.28 Semantic Memory - web.colby.eduweb.colby.edu/memoryandlanguagelab/files/2012/07/Balota-Coane-2008.pdf2.28.5 Grounding Semantics 520 2.28.5.1 Grounding Semantics in Analyses of