REVIEWS

m e m o ry s y s t e m s

Where do you know what you know?

The representation of semantic
knowledge in the human brain
Karalyn Patterson*, Peter J. Nestor‡ and Timothy T. Rogers§
Abstract | Mr M, a patient with semantic dementia — a neurodegenerative disease that is
characterized by the gradual deterioration of semantic memory — was being driven through
the countryside to visit a friend and was able to remind his wife where to turn along the not-
recently-travelled route. Then, pointing at the sheep in the field, he asked her “What are
those things?” Prior to the onset of symptoms in his late 40s, this man had normal semantic
memory. What has gone wrong in his brain to produce this dramatic and selective erosion of
conceptual knowledge?

Semantic memory (also called conceptual knowl- A semantic hub?

edge) is the aspect of human memory that cor- This Review focuses on one specific issue. Essentially all
responds to general knowledge of objects, word current theoretical positions about semantic memory
meanings, facts and people, without connection to share the view that much of the content of our semantic
any particular time or place 1. Knowing that you memory relates to perception and action, and is repre-
ate Szechuan scallops at the Peking Restaurant in sented in brain regions that overlap with, or possibly
Cambridge last Thursday evening is episodic, not even correspond to, the regions that are responsible for
semantic, memory. Knowing that Szechuan refers perceiving and acting4–6. This view about the neural
to a province of China, that food from this region tends to representation of how objects look, sound, move and
be spicy and that scallops are sea-creatures that live in so on therefore entails commitment to the idea that
brittle bivalve shells are all forms of conceptual knowl- conceptual knowledge is a widely distributed neural
edge. Memory for episodic events is not only specific network. Our knowledge of the scallop, for example,
to times and places, it is also largely specific to an includes attributes such as its visual features, which will
individual. Conceptual knowledge, on the other hand, be represented in or near the brain regions that analyse
is mostly shared across individuals in a given culture, visual form and colour; its manner of moving on the
although its precise scope depends on the individual’s seabed, represented in or near the brain regions that
experience. respond to the perception of this kind of movement;
Only a few decades ago, one respectable position its texture and taste, involving tactile and gustatory
*MRC Cognition and Brain held that semantic memory might arise from ‘uni- regions; the actions its edibility affords, such as cutting
Sciences Unit, 15 Chaucer versal connectivity’ in the brain, and hence have no with a knife and fork or chewing, supported by frontal
Road, Cambridge,
CB2 7EF, UK.
corresponding stable neural architecture 2. Because and parietal areas; its name or other descriptions that
‡
Neurology Unit, Cambridge we now know that brain lesions can produce selec- we could apply to it, represented in perisylvian language
University, Cambridge, UK. tive disruption of semantic memory, this hypothesis regions; and so on. The debate centres on the following
§
Department of Psychology, is probably no longer tenable; however, localizing question: are these distributed brain regions, along with
1202 West Johnson Street,
the stable architecture of conceptual knowledge has the connections between them, the entire neural basis
University of Wisconsin-
Madison, Madison, proved difficult. Indeed, one researcher concluded of semantic memory?
Wisconsin, USA. that “The search for the neuroanatomical locus of Many theories, forming a class that we refer to as the
Correspondence to: K.P. semantic memory has simultaneously led us nowhere distributed-only view (FIG.1a) offer a positive answer to
and T.T.R. and everywhere...” (Ref. 3) It is perhaps time for an this question. An alternative position, for which we
e-mails:
karalyn.patterson@mrc-cbu.
assessment of what we do and do not know about argue here, is that the sensory‑, motor- and language-
cam.ac.uk; [email protected] the representation of conceptual knowledge in the specific aspects of conceptual knowledge are necessary
doi:10.1038/nrn2277 human brain. but not sufficient: this is the distributed-plus-hub view

976 | december 2007 | volume 8 www.nature.com/reviews/neuro

© 2007 Nature Publishing Group
 REVIEWS

a Distributed-only view Gating architecture The distributed-plus-hub view is by no means the

Action Sound
first to argue for unified conceptual representations
Name Action Colours Motion that abstract away from modality-specific attributes7,8.
Motion
Most earlier proposals of this nature, however, were
mute regarding the neuroanatomical basis for this
central aspect of semantic processing. With a more
specifically neuroanatomical focus, Damasio and
colleagues9–13 proposed the existence of ‘convergence
zones’ that associate different aspects of knowledge,
Task
and along with other researchers14, they have clearly
Words Shape Task- articulated the importance of such zones for semantic
dependent processing. The convergence-zone hypothesis, however,
Shape Colour representation differs in at least two respects from the distributed-plus-
hub view illustrated in FIG. 1b. First, it proposes the
existence of multiple specialized convergence regions:
b Distributed-plus-hub view Convergent architecture for example, one that encodes associations between
Name Action Colours Motion visual representations of shape and corresponding
actions, another that encodes associations between
shape and object name, and so on. Second, it suggests
Task- that these zones become differentially important for
dependent representing different semantic categories. For example,
representation because humans frequently interact with tools and other
Task- man-made objects, the zone that links object shape
independent and action might be more important for knowledge of
representation
Task man-made artefacts than for knowledge of living things.
Similarly, because animals move in characteristic ways,
Shape the zone that links shape to movement might acquire
Figure 1 | Two theoretical positions regarding the neuroanatomical distribution special salience for knowledge of animals. These two
of the cortical semantic network and schematic models based on these views. aspects of the convergence-zone hypothesis make it a
Nature Reviews | Neuroscience variant of the distributed-only class of theories that is
Both positions hold, in agreement with most investigators, that the network is widely
distributed and partly organized to conform to the neuroanatomy of sensory, motor and represented in FIG. 1a. By contrast, the distributed-plus-
linguistic systems. a | The distributed-only view suggests that these widely distributed hub view proposes that, in addition to direct neuro-
regions, along with the diverse connections between them (shown as green lines), anatomical pathways between different sensory, motor
constitute the whole semantic network. The flow of activation through this network can and linguistic regions, the neural network for semantic
be ‘gated’ by a representation of the current task (right-hand panel): for instance, if the
memory requires a single convergence zone or hub that
task is to name a line drawing of a familiar object, activation will flow from a
supports the interactive activation of representations in
representation of object’s shape to a representation of its name. Associations between
different pairs of attributes are encoded along different neuroanatomical pathways. all modalities, for all semantic categories.
b | By contrast, the distributed-plus-hub view posits that, in addition to these modality- Do the distributed-only and distributed-plus-hub
specific regions and connections, the various different surface representations (such as positions result in different predictions? From the dis-
shape) connect to (shown as red lines), and communicate through, a shared, amodal ‘hub’ tributed-plus-hub perspective, damage to the hub should
(shown as a red area) in the anterior temporal lobes. At the hub stage, therefore, produce a semantic impairment that is independent of the
associations between different pairs of attributes (such as shape and name, shape and modality of input (objects, pictures, words, sounds, tastes,
action, or shape and colour) are all processed by a common set of neurons and synapses, and so on) and of the modality of output (for example,
regardless of the task. The right-hand panel (labelled ‘convergent architecture’) naming an object, drawing it or using it correctly). By
illustrates the model equivalent of the distributed-plus-hub view.
contrast, from the distributed-only perspective, no form
of focal brain damage would be expected to engender
such a generalized impairment. We therefore start with
(FIG.1b).The principal reason for this claim is that a two important empirical questions. Do selective but gen-
central function of semantic memory is to generalize eralized impairments of semantic memory occur? And,
across concepts that have similar semantic signifi- if so, are they caused by relatively focal brain damage?
cance but not necessarily similar specific attributes. The evidence pertinent to these questions comes from
Scallops and prawns have different shapes, colours, studies that have both investigated the performance of
shell structures, forms of movement, tastes, names, semantically impaired patients and attempted to deter-
verbal descriptions and so on, but semantically speak- mine the specific locus of their brain lesions. Following
ing, to seafood-eating humans they enter into similar an overview of this evidence, we ask to what extent the
scenarios and have substantial conceptual overlap. resulting conclusions are consistent with, or challenged
If semantic memory consisted only of the modality- by, evidence from functional neuroimaging studies of
specific content of objects (and the links between healthy adults performing semantic tasks. Finally, we
them), it is doubtful that we could ever achieve the consider evidence from computational modelling that
higher-order generalizations on which so much of our might help to explain why the cortical semantic network
semantic processing relies. is apparently organized in a particular way.

nature reviews | neuroscience volume 8 | december 2007 | 977

© 2007 Nature Publishing Group
REVIEWS

Box 1 | The syndrome of semantic dementia (ATL) is invariably evident by the time that SD patients
become symptomatic. Bilateral degeneration is the norm,
One of the most prominent symptoms of semantic dementia (SD) is anomia91: failure to although it is usually asymmetrical: in approximately
name objects, concepts and people, whether in response to stimulus presentation or in two-thirds of cases the atrophy is greater on the left side
spontaneous speech. Patients with other causes of anomia, and occasionally even
than on the right. The typical pattern seen on structural
healthy people, sometimes fail to name something because, although they know what
MRI scans is one of well-defined atrophy of both the left
it is, they cannot find the word for it at that moment. Anomia in SD does not result from
this kind of word-finding difficulty but instead reflects degraded knowledge of the and the right ATL, which is maximal at the temporal
object or concept. When we asked one of our patients to name a picture of a zebra, she pole and on the adjacent rostral-inferior surface.
replied: “It’s a horse, ain’t it?” Then, pointing to the stripes, she added, “But what are
these funny things for?” Alzheimer’s disease. The most prominent cognitive
For SD patients, the degree of success in naming (and, indeed, in any measure of deficit in typical AD is an impairment of episodic
conceptual knowledge) is largely independent of the modalities of input and output. (autobiographical) memory: in particular, the ability to
Naming is most commonly assessed by presenting pictures and asking for a verbal learn new information is progressively abolished. The
naming response; however, SD patients are just as unsuccessful, indeed usually more original hypothesis, which attributed this phenomenon
so, if the stimulus is a description of the item to be named (for example, “What do we
specifically to degeneration of the hippocampus20, is
call the African animal with black and white stripes?”), or is the item’s characteristic
now viewed not as wrong but as incomplete: functional
sound (for example, a telephone ringing or a dog barking), or if the patient is asked to
write the name rather than say it. brain imaging, even in patients with mild or early AD,
The degree of success or failure in any semantic test in SD patients is determined by reveals hypometabolism not only in the bilateral medial
four principal factors72,92: the severity or stage of progression of the disease; the temporal lobes, but also in the thalamus, the posterior
familiarity of the object (the deficit is always more pronounced for less familiar things, cingulate gyrus and other parts of the limbic system,
words and concepts); the object’s typicality (there are fewer successes for things that which apparently constitute a network that is crucial for
are atypical of their kind); and the specificity of information required by the task (for the formation of new memories21. Semantic memory is
example, a patient might be able to identify an apple as something to eat without frequently affected in AD22,23, but typically at a later stage
being able to identify it as an apple). and to a more modest extent than episodic memory.
Most cognitive functions apart from semantic memory are reasonably well-preserved
There are a few reported cases in which loss of concep-
in patients with SD, at least until late in the disease. Thus — provided that the tests do
tual knowledge was an early and significant symptom
not refer to or require access to semantic knowledge — SD patients have reasonably
normal capabilities in general intelligence, problem solving, visuo-spatial function, of AD. Where detailed neuroanatomical information
short-term memory, episodic memory, simple calculation skills, and so on. was available in these cases, however, the pathology
sometimes had a distribution that was atypical for AD
and was in fact more similar to that seen in SD24. It is
the contrast between the prototypical distributions of
Acquired disorders of semantic memory degeneration in these two diseases that is most germane
Impaired conceptual knowledge is associated with four to the current Review, because the relatively widespread
principal neurological aetiologies: stroke; viral infec- degeneration that is observed in AD apparently leads to
tion, most commonly herpes simplex virus encephalitis a less consistent and usually less profound disruption
(HSVE); and two forms of neurodegenerative disease: of semantic memory than is observed in SD, where the
Alzheimer’s disease (AD) and the semantic variant of pathology is relatively circumscribed within anterior
fronto-temporal dementia, typically called semantic temporal regions.
dementia (SD). All four conditions give useful clues about
the nature and organization of conceptual knowledge, Herpes simplex virus encephalitis. The most prominent
but from different angles. cognitive consequences of typical HSVE are amnesia for
earlier episodes in one’s life and profound difficulty in
Semantic dementia. SD is one of the neurodegenerative learning new information. The impact on episodic mem-
conditions that belongs to the Fronto-Temporal Dementia ory can be attributed to the fact that this viral infection
spectrum15. Recent post-mortem analyses indicate that has a predilection for medial temporal lobe structures,
the typical neuropathology in SD is abnormal neuronal including the hippocampus. Some patients with HSVE
inclusions of the protein ubiquitin16, a form of pathol- have no significant disruption to their conceptual knowl-
ogy that is best known in motor neuron disease. SD is edge, but have other cognitive deficits, most notably in the
marked behaviourally by a progressive deterioration executive and control abilities that are associated with
of expressive and receptive vocabulary and of knowledge the frontal lobe. When semantic memory is affected, the
about the properties of everyday objects, in the context deficit is often mild relative to the profound semantic def-
of otherwise well-preserved cognition and memory for icits that are observed in SD. As in SD, however, semantic
Expressive vocabulary recent events (BOX 1). It was first reported in the modern impairment in HSVE is associated with bilateral ATL
The set of words that an
era by Warrington17, Schwartz et al.18 and Snowden et al.19. lesions25,26. Interestingly, the semantic deficit in HSVE
individual knows and can
retrieve for referring to objects As indicated in BOX 1, research on SD provides clear is often category-specific, with relatively well-preserved
and other concepts in speech and positive answers to the two questions posed above. knowledge of man-made things but impaired knowledge
or writing. That is, SD represents a selective impairment to seman- of living things. There is considerable debate in the lit-
tic abilities that affects all modalities of reception and erature as to whether these two domains of conceptual
Receptive vocabulary
The set of words that an
expression, for all kinds of concepts, more or less equally knowledge are represented separately in the brain27, or
individual can comprehend (FIG. 2), and it is the consequence of relatively focal brain whether this dissociation can be explained by some fun-
when hearing or reading them. lesions. Degeneration of the anterior temporal lobes damental difference in the nature of the attributes that are

978 | december 2007 | volume 8 www.nature.com/reviews/neuro

© 2007 Nature Publishing Group
 REVIEWS

a Picture categorization b Picture naming

3 General Item Sept 1991 Mar 1992 Sept 1992 Mar 1993
Basic
Bird + + + Animal
Specific
Chicken + + Bird Animal
2 Duck + Bird Bird Dog

Accuracy (d′)
Swan + Bird Bird Animal
Eagle Duck Bird Bird Horse
1 Ostrich Swan Bird Cat Animal
Peacock Duck Bird Cat Vehicle
Penguin Duck Bird Cat Part of animal
Rooster Chicken Chicken Bird Dog
0
Controls Mild SD Moderate SD Severe SD

c Colour and object recognition d Delayed-copy drawing

Task one stimulus: Which is coloured correctly? Task one performance Model Delayed copy
Target-typical Target-unusual

1.00
Proportion correct
0.75

0.50

0.25

0.00
Controls Mild SD Severe SD

Task two stimulus: Which is the real animal? Task two performance
1.00
Proportion correct

0.75

0.50

0.25

0.00
Controls Mild SD Severe SD
Figure 2 | Examples of impaired performance on semantic tasks in patients with semantic dementia. This figure
illustrates the cross-modal nature of the impairment and the preservation of general relativeNature Reviews
to specific | Neuroscience
information in
semantic dementia (SD). a | Accuracy at discriminating targets from distractors in a picture-categorization task for four
groups of participants: healthy controls and patients with mild, moderate and severe SD72. Participants viewed a
category label followed by a colour photograph and were asked whether the picture matched the label. Labels were
either general (for example, ‘animal’), basic-level (for example, ‘dog’) or specific (for example, ‘Labrador’). b | Picture-
naming responses for one SD patient who was assessed longitudinally68. + denotes a correct response. c | Examples of
stimuli and performance (for controls, mild and severe SD patients) on two recognition tasks. In the first task,
participants judged which of two items was coloured correctly. Patients with milder SD performed well when the
targets had a category-typical colour (for example, the green celery) but poorly when the items had an unusual colour
(for example, the orange pumpkin). The judgements of patients with more severe SD were no better than chance (50%)
in either condition80. In the second task, participants judged which of two drawings depicted a real animal. Here, both
the patients with mild SD and the patients with severe SD achieved normal levels of success for targets with relatively
prototypical features (for example, the monkey which, like most animals, has small ears). For stimulus pairs in which the
correct choice had unusual features (for example, the elephant, which has very large ears), the patients with mild SD
were impaired and the patients with severe SD scored at chance levels70. d | Delayed-copy drawings produced by
SD patients71. The patients were shown a model picture which was then removed and, after a 10-second delay, they
were asked to reproduce this picture from memory. Properties that are common to most animals, such as eyes and a
tail, were preserved in the delayed drawings. Unusual properties that distinguish one animal from others — for
example, the hump on the camel and the flippers on the seal — were frequently omitted. Some common properties
were also incorrectly added to animals that lack them (for example, the four legs on the delayed drawing of the duck
and the tail on the delayed drawing of the frog). Real animal pictures in part c reproduced, with permission, from REF. 97
 (1980) American Psychological Association.

nature reviews | neuroscience volume 8 | december 2007 | 979

© 2007 Nature Publishing Group
REVIEWS

Aphasia characteristic of man-made and living things (for exam- The suggestion that the cognitive syndrome in SD
Impaired language abilities ple, living things might be more heavily weighted towards might result in part from structural and/or functional
resulting from brain disease or sensory features and man-made artefacts might be more abnormalities elsewhere in the semantic network 5
injury. heavily weighted toward functional attributes)28,29. Once is challenged by recent findings with (18F)fluoro‑2-
Cross-modal
again, the contrast between HSVE and SD in terms of deoxy‑D-glucose positron emission tomography
The term that is applied to both the severity of the semantic deficit (which is often (FDG-PET). Three studies37–39 have confirmed bilateral
representations or processes either absent or mild in patients with HSVE, as opposed anterior temporal dysfunction in SD, but no other area
that operate across different to progressive and ultimately profound in SD) and its of hypometabolism was consistently observed: one
kinds of sensory, motor and
pattern (which is frequently category-specific in HSVE, reported more extensive hypometabolism along the
linguistic representations. For
instance, representations and
but very rarely so in SD) seems most telling in the current length of the inferior left temporal lobe37, the second
processes that receive input context. That is, because both diseases implicate the bilat- observed additional hypometabolism in the left insula
from and/or direct output to eral ATL in semantic processing, it must be the specific and orbito-frontal areas38 and the third revealed no sig-
both visual and auditory nature and/or distribution of the brain abnormalities in nificant reduction in metabolism in any region except the
representations would be
considered cross-modal.
SD that produces the pervasive disruption — across all rostral temporal lobes39. In agreement with these studies
categories and all modalities — of conceptual knowledge of resting metabolism, volumetric structural imaging40
Anomia that defines this condition. indicates relative preservation of the posterior temporal
The failure to name objects, lobe in SD38,39,41–43. In the dorsal-ventral plane, the supe-
concepts and people, whether
Stroke. The most prominent impairment in stroke rior temporal gyrus is atrophic but also preserved rela-
in response to stimulus
presentation or in spontaneous
patients with extensive lesions in the left hemisphere is tive to the inferior temporal gyrus40. In other words, both
speech. aphasia, but poor performance on non-verbal semantic structural and metabolic imaging studies of patients with
tests as well as in verbal comprehension can result from SD suggest that the abnormalities are most pronounced
Volumetric MRI left-hemisphere stroke, especially in a condition called in the anterior and inferior parts of the temporal lobes.
A method that uses finely cut
brain slices (usually less than
transcortical sensory aphasia (TSA). At least in this Of course, it is possible that the pathology is more
2 mm thick) to measure the cross-modal regard, TSA might resemble SD; however, widely distributed in SD, but that such further abnor-
volume of brain structures. there are important differences between the two patients malities are so subtle or variable across patients that
groups. With reference to criteria designed to distinguish they do not meet the standard for statistical significance
Voxel-by-voxel analysis
impairments of representation from problems in access in voxel-by-voxel analyses. A recent comparison of brain
A method of whole-brain image
analysis in which the brain
or retrieval30, some recent work suggests that the seman- hypometabolism in patients with AD versus patients
scans of different individuals tic deficit in TSA may be better described as an impaired with SD suggests that this is not the case. Nestor and
are fitted to a standard ability to retrieve, select and manipulate semantic infor- colleagues39 demonstrated that hypometabolism is much
template (to minimize inter- mation in a task-appropriate fashion31, rather than the more widespread in AD than in SD and, in particular,
individual differences in brain
shape) so that brain regions
degradation of semantic representations themselves that it affects left frontal, occipito-temporal and tem-
can be compared which is characteristic of SD. For example, the anomia poro-parietal regions in AD — areas that have been
systematically across subjects. that occurs in TSA readily benefits from cueing (for a implicated in the distributed cortical semantic network
patient struggling to name a picture of a violin, the cue and which appear relatively normal in SD. Yet semantic
Lesion-overlap study
might be, “It begins with a ‘v’”)32, whereas anomia in SD impairment was much milder in AD patients than in
A method that seeks to define
a common area of brain
is largely unaided by cueing32,33. There is essentially no SD patients, all of whom showed hypometabolism in the
damage relevant to a given overlap between the brain regions that are damaged in ATL (FIG. 3).
behavioural deficit by these two patient populations. Owing to the anatomy of Other lesion studies also support the view that the
overlaying the scan-defined the vascular system, stroke rarely, if ever, produces focal ATL is important for semantic representation and/or
lesions of multiple subjects
with the behavioural deficit in
lesions in the ATL, and semantic deficits in TSA typically processing. Perhaps the best known is a lesion-overlap
question. result from damage to either frontal or parietal regions study that tested picture naming in a large group of
(or both) that is restricted to the left hemisphere34. In anomic patients with unilateral, stable, focal brain
PET activation paradigm accordance with this view regarding the nature of the lesions11. The stimuli included pictures of famous peo-
An experimental paradigm that
semantic impairment in TSA, there is substantial evi- ple — who must, of course, be named at a specific level
uses PET to measure changes
in cerebral perfusion in
dence from functional imaging of normal individuals (for example, ‘Princess Diana’ rather than ‘princess’ or
response to a stimulus. that activation of prefrontal cortical areas is associated ‘woman’) — and pictures of animals and tools, which are
with selection or control processes35,36. usually named at the basic level (for example, ‘elephant’
or ‘hammer’), which does not distinguish individual
Summary of semantic disorders. The range of aetiologies category members from one another. Impaired naming
associated with impaired performance on semantic tasks of famous faces was associated with the ‘tightest’ lesion
reveals important similarities and differences. The cogni- overlap, centred on the left temporal pole, whereas
tive and neuroanatomical abnormalities in SD, especially impaired animal naming correlated with lesions in the
when contrasted with those from the other aetiologies anterior, inferior left temporal lobe. Poor tool-naming
reviewed above, provide trenchant evidence for the revealed the lowest degree of lesion overlap and was
distributed-plus-hub view (FIG. 1b). The strength of this associated with damage in the posterior, lateral left tem-
evidence hinges, however, on claims about the relatively poral lobe as well as in the temporo–occipito–parietal
focal nature of the pathology in SD. The remaining issue junction. Lesion–symptom correlation thus implicated
to be considered in this section is whether the selective the left ATL in naming for two of the three categories
conceptual impairment in SD can be securely attributed tested. When healthy participants named the same
to circumscribed lesions of the ATL. items in a PET activation paradigm, all three stimulus

980 | december 2007 | volume 8 www.nature.com/reviews/neuro

© 2007 Nature Publishing Group
 REVIEWS

a Semantic dementia Alzheimer’s disease Evidence from functional neuroimaging

This conclusion from neuropsychological evidence
might seem surprising to researchers who are famil-
iar with functional brain-imaging studies of semantic
abilities in healthy individuals. Although there have
been more than 40 published reports of ATL activation
by a wide variety of semantic tasks (FIG. 4), these reports
constitute a relatively small proportion of the functional
imaging studies of semantic memory in the literature of
the past decade. According to both primary sources44–49
and review articles3,5,50,51, imaging studies more often
implicate some combination of frontal, posterior tempo-
ral, temporo-parietal and parietal regions in the cortical
b semantic network. In this section we consider whether
SD AD Control these results are genuinely at odds with the neuro­
70 psychological data discussed above, and suggest two
1.0 ways of reconciling the apparently conflicting sources
60 of evidence.

50 0.75 The anterior temporal lobe is ‘shy’ to functional MRI.

Proportion correct

In standard acquisition protocols for functional MRI

40
(fMRI), which has largely replaced H215O-PET for func-
Score

0.5 tional neuroimaging purposes, the signal-to-noise ratio

30
diminishes substantially near the temporal poles, owing
20
to their proximity to air-filled sinuses (the so-called
0.25 ‘susceptibility artefact’). One study compared the func-
10 tional activation revealed by PET and by fMRI (using a
standard whole-brain acquisition sequence) in a group
0 0.0 of people performing a semantic categorization task52.
Category fluency PPTp PPTw Naming WPM Although both methods showed substantial semantic-
Figure 3 | Differences between semantic dementia and Alzheimer’s disease in related activity in ventral posterior temporal lobes, only
measures of brain function and semantic memory. a | The areas of reduced
Nature Reviews | Neuroscience
PET revealed that this activation extended all the way
metabolism (shown as graded grey areas), are widespread in patients with Alzheimer’s rostrally to the temporal pole. Other PET evidence has
disease (AD) and include some regions that are implicated in the cortical semantic demonstrated significant ATL activation during the
network (see FIG. 1). In the AD cases shown, however, there was little evidence of any semantic tasks of category fluency53, object naming54,
abnormality in anterior temporal regions, which show substantial and focal category verification55 and word recognition56. Perhaps
hypometabolism in patients with semantic dementia (SD). b | The performance of AD
most saliently — because what could more clearly exem-
and SD patients, as well as of a group of age-matched healthy controls, on a range of
semantic tasks. The SD patients were significantly more impaired than the AD patients plify the essence of semantic processing? — the tasks of
on all tasks, even though the brain abnormalities in the AD patients were more comprehending connected speech versus various forms
widespread. In category-fluency tasks, participants are given 1 minute to list as many of distorted (and hence meaningless) speech, and of
examples of a semantic category as they can. PPTp and PPTw indicate the picture and reading coherent text versus a meaningless visual control
word versions, respectively, of the Pyramids and Palm Trees test of semantic association98. condition, seem to engage the left temporal pole when
Naming indicates performance on a simple picture-naming task and WPM indicates measured by both PET57,58 and fMRI59.
performance on a 10-alternative, forced-choice word-to-picture matching task. Other functional imaging methods similarly impli-
Parts a and b modified, with permission, from REF. 39  (2006) Academic Press. cate the ATL in semantic processing. For instance, an
elegant experiment was performed with anatomically
constrained magnetoencephalography (MEG) in which
types yielded significant blood-flow increases (relative subjects made semantic judgements about spoken or
to a control condition) in the left temporal pole, with written words60. As expected, the brain’s initial responses
further activation of the right temporal pole during to words were detected in the appropriate sensory areas
face-naming and similar but discernibly different acti- for each modality; however, from approximately 400
vation patterns in the posterior left temporal regions for ms after stimulus onset, the MEG responses for both
animal- versus tool-naming11. modalities converged on the ATL.
Taken together, the neuropsychological literature
Magnetoencephalography suggests that the selective but amodal semantic impair- Specificity of semantic processing. In the PET component
(MEG). A method of measuring ment in patients with SD is attributable to relatively focal of the object-naming study cited above11, the strong-
physiological activity across pathology in the ATL and not to widespread damage est ATL activation was observed when participants
the cortex by detecting in the cortical semantic network. Such data strongly were required to recognize and name famous people.
pertubations in the magnetic
field that is generated by the
support the distributed-plus-hub view shown in FIG. 1b, A subsequent study from the same group61 proposed
electrical activity of neuronal and further suggest that the ‘hub’ part of the semantic that the temporal pole might have a special role in
populations. network is located in the bilateral ATL (BOX 2). the recognition and identification of unique concepts

nature reviews | neuroscience volume 8 | december 2007 | 981

© 2007 Nature Publishing Group
REVIEWS

Box 2 | Why the anterior temporal lobe?

This Review summarizes evidence for the hypotheses that, first, semantic generalization requires a single amodal hub
and, second, that the neuroanatomical site of this hub is the anterior temporal lobe (ATL). Assuming, for the moment, that
these hypotheses are valid, why is the ATL a neuroanatomically sensible place for the semantic hub? One reason that is
sometimes offered is the fact, known from non-human primate physiology93, that many primary sensory and motor areas,
along with their related association cortices, connect to the ATL. This is true, but not a unique feature of the ATL: the ‘tri-
state’ junction of the temporal, parietal and occipital lobes, including the angular gyrus, has the same feature of being
well connected to input from multiple modalities, and indeed some researchers have ascribed a similar cross-modal
mapping function to this area14,94. It is worth noting that cross-modal is not the same as a‑modal: the region around the
angular gyrus might serve to combine information from several modalities but still not have the genuinely amodal
function of a semantic hub.
Two neuroanatomical facts about the ATL region might make it an appropriate candidate for extracting amodal
conceptual information. First, it is neuroanatomically proximal to the amygdala and to other limbic structures, as well
as to the orbito-frontal cortex — regions that are known to be important for the processing of emotion and reward95.
Given that affective response to some extent pervades everything that we perceive, do and know, ATL regions might be
well suited to computing associations between affect and more value-neutral sensory, motor and linguistic aspects of
conceptual knowledge. Second, the ATL regions that seem so crucial to semantic memory are immediately adjacent to
the anterior parts of the medial temporal lobe memory system — a system that is critical for rapid learning of new
episodic information. As episodes must contribute to the gradual acquisition of new conceptual knowledge, it makes
sense for episodic and semantic systems to be situated in close proximity.

— individual people and famous buildings being two Functional imaging studies may likewise indicate
prominent examples. A number of functional imaging that the ATL is most strongly recruited not just by rec-
studies, using both PET and fMRI, have since supported ognition of unique items (such as the Eiffel Tower or
this idea, showing ATL activation related to the recogni- Princess Diana), but by any form of specific semantic
tion of familiar versus unfamiliar buildings62, faces62,63, processing. One of several PET studies that support
names64,65 and even voices66. this view used a task in which normal participants
Semantic dementia patients are profoundly impaired were asked to verify, by answering yes or no, whether
at recognizing famous individuals from photographs, the stimulus shown in a colour photograph belonged to
names and verbal descriptions67; this impairment, a particular category55. For example, in trials designed
however, appears to be one notable manifestation of a to receive ‘yes’ responses, participants judged on dif-
more general sensitivity to the specificity with which ferent occasions whether the photo of a robin depicted
an item must be recognized or categorized. Across a an animal (superordinate level), a bird (basic level) or
wide array of semantic tasks (FIG. 2), the patients might a robin (subordinate or most specific level). Relative to
perform well as long as accurate performance requires a control condition, all three semantic-judgement con-
only a relatively coarse or general categorization of ditions activated the bilateral posterior fusiform gyrus
the stimulus. Thus, severely impaired SD patients can and the occipito-temporal cortex, but the subordinate
sometimes call a picture an ‘animal’, without being able condition, when contrasted with both superordinate
to name it ‘chicken’ or even ‘bird’68; they can accurately and basic levels, activated the ATL bilaterally. The
sort pictures or words into categories such as ‘animal’ ATL activation peaks aligned closely with the areas
versus ‘man-made object’, but not ‘car’ versus ‘boat’69; that showed the strongest grey-matter reduction in
they will correctly judge — when offered two pictures a voxel-based morphometry analysis of atrophy in SD
of a donkey, one with and one without a hump on its patients42. The ATL activation peaks for the subordi-
back — that ‘the real one’ is the humpless exemplar, but nate condition also matched those from another PET
will make the same humpless (and therefore more ‘ani- study that contrasted the naming of unique items with
mal prototypical’, but incorrect) choice for the camel, to the (basic-level) naming of common objects73. Similar
which humps are specific70; they will copy-draw a line results have been reported using fMRI74,75. A conjunc-
drawing of a real humped camel a mere 10–15 seconds tion analysis of results from four different PET studies
after studying it, but omit the hump71; and so on. This of semantic processing, two with words as stimuli and
Voxel-based morphometry
A voxel-by-voxel analysis of
pattern does not arise simply because tasks that require two with pictures, highlighted activation of the inferior
structural-image data, most precise classification are more difficult. For example, ATL; the authors concluded that this region supports a
commonly the grey-matter Rogers and Patterson72 showed that, whereas healthy polymodal or amodal network of semantic representa-
segments extracted from controls are faster and more accurate at classifying tions that is recruited when more specific conceptual
T1‑weighted MRI.
items at the basic level (for example, ‘dog’) relative to information is required56 (BOX 3).
Conjunction analysis a more general level (for example, ‘animal’), patients This brief Review suggests that activation in ATL
A statistical method used in with SD show the reverse pattern: they have greater regions during semantic tasks is not as scarce as an initial
functional brain imaging impairment on the more precise basic-level classifica- survey of the literature might suggest, especially if one
research for identifying brain tion. Such findings suggest that semantic tasks that takes into account the low signal-to-noise ratio that is
regions that are significantly
activated in two or more
require the distinctive classification of a stimulus place achieved by standard fMRI methods in these regions
separate experimental particularly strong demands on the ATL regions that and the fact that the amount of signal generated in ATL
conditions. are affected in SD. regions is related to the specificity of semantic process-

982 | december 2007 | volume 8 www.nature.com/reviews/neuro

© 2007 Nature Publishing Group
 REVIEWS

Unique recognition Basic-level naming Other basic-level tasks

63 105
64 66 107
55 54 56 56
55 65 54 52
65 64 23
62 62 104

65
52
64 104
55 54
107 56 23
62 105

64
65
62
55
54 56
63
66

Speech comprehension Reading/word recognition Other semantic tasks

106 112
57 57 104 106 53
100 109
58 104 113 106
100 99 108 102
58 103 111 110
53
Non-linguistic
Linguistic

Both

Auditory 100

109 106
Visual 108
104 103 53
58 113
57 112
Both 106 111

Other 106

58
99 102
Generation task 100 104
53
57 110
Figure 4 | ATL activation in functional imaging studies of semantic processing. This figure shows the most anteriorly
located temporal-lobe activation peaks (plotted in Talairach space) reported in 28 differentNature
studies of semantic-task
Reviews | Neuroscience
performance in healthy individuals. More than 40 of such studies using a variety of imaging methods, have been
published; this set includes those studies that used either positron emission tomography or functional MRI and
reported coordinates of anterior temporal lobe (ATL) peaks in either Montreal Neurological Institute (MNI) or Talairach
standardized spaces. Peaks originally reported in MNI coordinates were transformed to Talairach space using the
transformation described by Brett and colleagues114. The studies varied in the modality of stimulus presentation, the
nature of the particular task that was being performed, whether or not the stimuli were linguistic, and the specificity of
the categorization that was required for successful task performance. Despite these variations in methodology, the
different studies showed activation in remarkably similar regions of the ATL. The numbers indicate the studies from
which the data were taken, as numbered in the reference list (see REFS 52–58,61–66,99–113). The shape and colour of
the plotted areas indicate the nature of the stimulus that was used in the experiments, as shown in the legend. Where
multiple points are plotted for the same study, these represent the most anteriorly located activation peaks in the
temporal lobe from different contrast conditions in the same study. For instance, for a study that used both words and
pictures as stimuli, and that reported separate anterior-temporal peaks corresponding to the two stimulus sets, both
peaks are shown.

nature reviews | neuroscience volume 8 | december 2007 | 983

© 2007 Nature Publishing Group
REVIEWS

ing demanded by the task. Of course, a review of the the posterior-temporal cortex and motor plans in frontal
functional imaging literature performed with a different and parietal areas, perhaps via the dorsal visual process-
‘filter’ might reach a different conclusion, especially if it ing stream? Yet the ATL focus of damage in SD, which
excluded PET studies and included only fMRI studies is far removed from these sites and does not lie between
that lacked techniques designed to maximize the signal in them, seems to disrupt knowledge about characteristic
the ‘hard-to-get’ rostral temporal lobes. Indeed, much of object use78,79. Furthermore, damage in the same region
the evidence from such imaging studies supports the view impairs knowledge of the colours that are character-
that semantic processing involves mainly frontal or pos- istic of certain object shapes (for example, grey for an
terior-temporal loci, and of course our own position also elephant and orange for a carrot)80,81, even though shape
implicates these regions in the larger semantic network. and colour are two manifestly visual forms of informa-
The current Review is selective in that our primary tion that are thought to be supported by neighbouring
aim was to determine whether the scarcity of reported regions in the posterior ventral occipito-temporal cortex.
ATL activation in the literature to date is sufficient to Again, why are there not just direct connections between
refute the conclusion drawn from lesion studies, which these two modality-specific regions? That is, why the
unambiguously point to ATL regions as being crucial for need for a single hub?
semantic processing. To paraphrase a recently-departed The second question relates to the fact that was noted
United States Secretary of Defense, we do not believe earlier regarding scallops and prawns: conceptual simi-
that the seeming absence of evidence constitutes, in this larities between items are not necessarily apparent from
case, evidence of absence. their perceptual features. For instance, in one of the sim-
ple tasks used in studies with healthy and semantically
Why does the semantic network need a hub? impaired individuals, called ‘category fluency’, a person
The above considerations support a view of human might be asked to name as many different fruits as possi-
semantic cognition that raises at least two questions. The ble in one minute. A typical sequence of responses might
first question is why a single region of the brain should be “Apple, orange, banana, pear, grapes, lemon…” and so
contribute to learned associations among widely distrib- on. These six objects are very different from one another
uted sensory, motor and linguistic representations. For in colour, shape, texture, how they grow, how they are
example, visual representations of objects are apparently eaten, et cetera. Yet normal individuals can (whereas SD
coded in the posterior inferior/middle temporal cortex, patients definitely cannot) perform this task with ease,
and knowledge about action is probably supported by because knowledge of conceptual similarity allows these
frontal and parietal regions76,77. Surely then, the knowl- objects to be grouped as fruits, even though they have
edge that allows one to brush one’s hair with a hairbrush few sensory/motor properties in common. The second
and one’s teeth with a toothbrush could be coded in question is therefore: how does the semantic system
direct associations between object-recognition regions in acquire representations that capture such conceptual
similarity relationships?

Box 3 | Sensitivity to specificity in the anterior temporal lobe Evidence from computational modelling. The two ques-
tions posed above might seem unrelated, but evidence
Both neuropsychological and functional imaging data suggest that anterior temporal from neural-network models suggests that they are not.
lobe (ATL) regions are especially taxed by tasks that require very specific recognition or Computer simulations with such models have shown
classification of a stimulus. What accounts for this sensitivity to specificity?
that networks that adopt a convergent architecture — in
One possibility is that a specificity gradient exists in the temporal lobes, such that the
greater the specificity with which an item must be recognized or categorized, the more which all forms of information about concepts are, at
rostral the activation in the temporal lobes will be96. The data in FIG. 4, however, some point, processed through the same population of
challenge this hypothesis: the ATL regions that are engaged by recognition of unique neurons and synapses — exhibit functional properties
entities (top left of FIG. 4) are no more anterior than those that are activated by tasks that explain how the semantic system is able to learn
that require discrimination of (non-specific) basic-level categories (top middle and top conceptual similarity relationships82. An example of such
right of FIG. 4) or other semantic tasks (bottom section of FIG. 4). a convergent architecture is illustrated in the right-hand
The distributed-plus-hub view offers a different explanation. It suggests that ATL panel of FIG. 1b. Contrasting with this is what might be
regions encode the similarity relations among various concepts, so that semantically called a ‘gating’ architecture, illustrated in the right-hand
related items (for example, various different birds) are coded with similar patterns across panel of FIG. 1a.
ATL neurons. According to this model, naming a particular bird as a ‘robin’ requires
In the gating architecture, the associations between
the ATL hub to instantiate the robin representation almost exactly, as the name does not
apply to other kinds of birds, many of which nevertheless have representations that are different kinds of attributes are encoded in different
very similar to the robin. To name the same item ‘bird’, however, the robin pattern need neuroanatomical pathways: one pathway stores the asso-
not be instantiated exactly. Because the name applies to all birds and all birds share ciation between an item’s shape and its name, another
similar representations, it is only necessary for the hub to find a representation that is stores the association between an item’s shape and the
sufficiently ‘bird-like’ to activate the name. Thus, small distortions of the ‘robin’ usual action associated with the item, and so on. A
representation — perhaps resulting from ATL atrophy — will prevent the network from representation of the current task determines the path-
retrieving the robin’s specific name (and other properties that differentiate it from other way through which activation will flow. If the task is to
birds) without disrupting the retrieval of properties that are common to birds. A similar name a line drawing of an object, then activation will
explanation extends to the interpretation of the functional imaging results, if one flow from the shape representation to the name repre-
assumes that a stronger metabolic response in the ATL occurs in tasks that require the
sentation. If the task is to demonstrate how an object
differentiation of highly overlapping representations.
is used, then activation will flow along a different path,

984 | december 2007 | volume 8 www.nature.com/reviews/neuro

© 2007 Nature Publishing Group
 REVIEWS

Connectionist model
from the shape representation to a representation of texture, and so on, as well as complementary mappings
A form of computational model action. In this architecture, the same input activates sepa- (that is, mappings in the other direction) between these
used to understand cognitive rate pathways, depending on the task at hand — thus, the surface representations, the internal representations that
processes by simulating the task representation ‘gates’ the flow of activation through emerge look very different. They are not dominated by
flow of activation among
simple, neuron-like processing
the network. the similarities expressed in any individual modality
units through weighted, By contrast, in the convergent architecture, the asso- (or pair of modalities), but instead reflect the similarity
synapse-like connections. ciations between a given input (for example, a shape relationships that is apparent across all of the modality-
representation) and all other kinds of attribute are stored specific representations taken together. In other words,
Normative property-listing in the same neuroanatomical pathway. As FIG. 1b the intermediate representations that arise in the hub can
studies illustrates, the task-dependent representation can still capture the ‘deep’ structure of concepts and hence
Studies in which participants
are given the name of a
shape the flow of activation through the network, so can promote generalization across items that are concep-
category or item (for example, that the same stimulus can differentially activate various tually related, even if they do not happen to have similar
‘bird’ or ‘robin’) and asked to attributes depending on the task demands. In this kind of shapes, colours, associated actions, and so on. These
list as many properties as they model, however, the neurons and synapses that comprise representations are amodal in that they can be generated
can think of that characterize
it. The proportion of
the task-independent representations contribute to the from any individual receptive modality and can be used
participants that list any processing regardless of whether the task is to name an to generate behaviour in any individual expressive modal-
particular property serves as object, use an object or identify an object’s characteristic ity. They are semantic in that they express the conceptual
an index of the property’s colour (when seen in black and white). similarity relations among concepts that are critical to
importance to the concept.
Computer simulations of learning and processing semantic generalization and induction, even though, in
Drawing-to-name
in networks with these architectures suggest that the themselves, they have no retrievable content82.
A non-verbal method of convergent architecture is better able to learn conceptual Rogers et al.88 demonstrated the appeal of this idea
investigating a person’s similarity relationships than the gating architecture83. To using a fully recurrent connectionist model that was trained
knowledge of objects in which understand this, consider how the gating model might to map between simple visual representations of objects,
the participant is given a
concept name (for example,
encode information about a pear. The pathway that verbal descriptions of the objects, and the objects’ names.
‘camel’) and asked to draw the stores associations between shape and name will learn The information contained in the verbal descriptions and
corresponding object. an intermediate representation that reflects both visual visual representations was derived from normative prop-
and phonological similarity to other known objects. erty-listing studies89 and a study of drawing-to-name88. The
Thus, a pear and a light-bulb will generate somewhat internal representations that were learned by the model
similar representations in this pathway because they captured the gross similarity relations among the items in
have similar shapes; a pear and a bear will generate the corpus. More interestingly, they also captured aspects
somewhat similar representations because they have of similarity structure that were not apparent when
similar-sounding names; and a pear and a banana will considering the verbal descriptions or the visual repre-
generate rather different representations because they sentations alone. For instance, considering just visual
have different shapes and different-sounding names. The similarities, fruits and vegetables share many properties
gating architecture will not encode conceptual similarity with man-made objects whereas, considering just the ver-
relationships, which should capture the fact that pears bal descriptions, fruits and vegetables are quite distinct
and bananas are semantically related whereas pears and from both animals and man-made objects, although they
light-bulbs or pears and bears are not. share a few properties with animals. When trained on
It might seem as though such problems can be solved these patterns, the model acquired internal representa-
simply by attributing greater weight or salience to some tions in which the fruits and vegetables were distinct from
sensory or motor features than others. For instance, if animals and man-made objects, but were actually more
similarity of taste is more salient than similarity of shape similar to man-made objects. Thus, the model made the
or word-sound, then bananas and pears, which are both counterintuitive suggestion that fruits and vegetables,
sweet, might be judged more similar to one another than although they are ‘natural’ and not man-made, might
pears and light-bulbs or pears and bears. The problem be represented as being more similar to artefacts than to
with this approach is that the salience of a given feature other natural things (like animals) in the human seman-
varies from one semantic category to another: colour, for tic system. Consistent with this suggestion, when asked
example, is important for categorizing fruits (consider to sort pictures of apples and other fruits and vegetables
lemons versus limes), but is irrelevant for categorizing into one of three categories — plant (correct), animal,
toys84. Thus, there appears to be a chicken-and-egg prob- or man-made artefact — SD patients mis-assigned a
lem: to determine the salience of a given sensory, motor number of fruits and vegetables to the artefact category,
or linguistic feature, one must know to which category despite making few errors when the choice categories for
the item belongs, but the item is difficult to categorize the apple were fruit, bird or land animal88.
without knowing the salience of its observed features85–87.
In other words, there is no single salience for a given Concluding remarks
property that will correctly capture semantic similarity We return to Mr M, driving through the countryside
for all concepts. with his wife, retrieving the several-years-old memory
In a convergent architecture, however, where the same that they will have to turn left ahead. His view from the
intermediate units that code the association between car window includes not just the cues to this preserved
shape and name must also learn to code relationships route knowledge (knowledge that many people with
between shape and colour, shape and action, shape and normal brains would find difficult to retrieve), but one

nature reviews | neuroscience volume 8 | december 2007 | 985

© 2007 Nature Publishing Group
REVIEWS

of the most familiar scenes in the British countryside: a with SD in dozens of neurology clinics in many differ-
flock of sheep. The sheep are a puzzle to Mr M: not only ent languages and countries. Not every such patient has
does he not know what to call them, he no longer knows had detailed structural and/or functional brain imaging
what they are. He wears a wool jacket when it’s cold and but, for those who have, the resulting neuroanatomi-
eats roast lamb for Sunday lunch, but would not be able cal profile is as consistent as the cognitive profile: the
to say that “those things” out there are the source of these selective but generalized semantic degradation that
products. He would succeed in matching a photograph occurs in SD goes hand-in-hand with focal degenera-
of a sheep taken from the side to one taken from the tion of the bilateral ATL. Specific features of conceptual
front, because this task — which people can perform knowledge are almost certainly represented elsewhere
on meaningless objects that they have never seen before and in a widely distributed network; but people’s ability
— relies on visual perceptual abilities rather than seman- to receive information in one modality and express it in
tic ones90. If asked whether the photograph of a sheep is another, to generalize across conceptually similar enti-
an animal, he would probably say yes, but if asked what ties that differ in almost every specific modality, and to
other animal is similar to it, he would look blank. This differentiate between entities that resemble each other in
striking combination of preserved and disrupted cogni- many modalities — all quintessentially semantic abilities
tion has now been documented in hundreds of patients — seem to depend on the ATL.

1. Tulving, E. in Organisation of Memory (eds Tulving, E. & 16. Davies, R. R. et al. The pathological basis of semantic 34. Berthier, M. Transcortical Aphasias (Psychology, Hove
Donaldson, W.) 381–403 (Academic, New York, 1972). dementia. Brain 128, 1962–1963 (2005). East Sussex, 1999).
2. Fodor, J. A. The Modularity of Mind (MIT Press, 17. Warrington, E. K. The selective impairment of 35. Badre, D., Poldrack, R., Paré-Blagoev E., Insler R. &
Cambridge, Massachusetts, USA, 1983). semantic memory. Q. J. Exp. Psychol. 27, 635–657 Wagner, A. Dissociable controlled retrieval and
3. Thompson-Schill, S. L. Neuroimaging studies of (1975). generalized selection mechanisms in ventrolateral
semantic memory: inferring “how” from “where”. This article represents one of the first, and still one of prefrontal cortex. Neuron 47, 907–918 (2005).
Neuropsychologia 41, 280–292 (2003). the very best and most creative, neuropsychological 36. Gold, B. T. et al. Dissociation of automatic and strategic
This review of neuropsychological and functional investigations of semantic dementia. lexical-semantics: functional magnetic resonance imaging
imaging studies offers a succinct overview of the 18. Schwartz, M. F., Martin, O. S. M. & Saffran, E. M. evidence for differing roles of multiple frontotemporal
modality-specific components of the cortical Dissociations of language function in dementia: a case regions. J. Neurosci. 26, 6523–6532 (2006).
semantic network and addresses the role of study. Brain Lang. 7, 277–306 (1979). 37. Diehl, B. et al. Cerebral metabolic patterns at early
prefrontal regions in control or selection of semantic 19. Snowden, J. S., Goulding, P. J. & Neary, D. Semantic stages of frontotemporal dementia and semantic
information. dementia: a form of circumscribed cerebral atrophy. dementia. A PET study. Neurobiol. Aging 25,
4. Allport, D. A. in Current Perspectives in Dysphasia Behav. Neurol. 2, 111–138 (1989). 1051–1056 (2004).
(eds Newman, S. K. & Epstein, R.) 32–60 (Churchill 20. Delacourte, A. et al. The biochemical pathway of 38. Desgranges, B. et al. Anatomical and functional
Livingston, Edinburgh, 1985). neurofibrillary degeneration in aging and Alzheimer’s alterations in semantic dementia: a voxel-based MRI
5. Martin, A. The representation of object concepts in disease. Neurology 52, 1158–1165 (1999). and PET study. Neurobiol. Aging 28,1904–1913.
the brain. Annu. Rev. Psychol. 58, 25–45 (2007). 21. Nestor, P. J., Fryer, T. D., Smielewski, P. & Hodges, J. R. 39. Nestor, P. J., Fryer T. D. & Hodges J. R. Declarative
This is an elegant review of evidence for a Limbic hypometabolism in Alzheimer’s disease and mild memory impairments in Alzheimer’s disease and
distributed network of semantic features linked to cognitive impairment. Ann. Neurol. 54, 343–351 semantic dementia. Neuroimage 30, 1010–1020
the sensory and motor systems that analyse such (2003). (2006).
features as input. It also argues for categorical 22. Giffard, B. et al. The nature of semantic memory deficits This study highlighted that the lesions (defined by
organization of semantic memory. in Alzheimer’s disease. Brain 124, 1522–1532 (2001). FDG-PET metabolism) associated with the semantic
6. Saffran, E. M. & Schwartz, M. F. in Attention and 23. Grossman, M. et al. Neural basis for semantic memory deficit in semantic dementia are restricted to the
Performance XV (eds Umilta, C. & Moscovitch, M.) difficulty in Alzheimer’s disease: an fMRI study. Brain anterior temporal lobes, whereas posterior temporal
507–535 (Lawrence Erlbaum Associates Inc., 126, 292–311 (2003). hypometabolism was evident in an Alzheimer group
Hillsdale, 1994). 24. Galton, C. J., Patterson, K., Xuereb, J. H. & Hodges, with absent or very mild semantic deficits.
7. Riddoch, M. J., Humphreys, G. W., Coltheart, M. & J. R. Atypical and typical presentations of Alzheimer’s 40. Chan, D. et al. Patterns of temporal lobe atrophy in
Funnell, E. Semantic systems or system? disease: a clinical, neuropsychological, neuroimaging semantic dementia and Alzheimer’s disease.
Neuropsychological evidence re-examined. Cogn. and pathological study of 13 cases. Brain 123, Ann. Neurol. 49, 433–442 (2001).
Neuropsychol. 5, 3–25 (1988). 484–498 (2000). 41. Davies, R. R., Graham, K. S., Xuereb, J. H., Williams,
8. Caramazza, A., Hillis, A. E., Rapp, B. C. & Romani, C. 25. Lambon Ralph, M. A., Lowe, C. & Rogers, T. T. Neural G. B. & Hodges, J. R. The human perirhinal cortex and
The multiple semantic hypothesis: multiple basis of category-specific deficits for living things: semantic memory. Eur. J. Neurosci. 20, 2441–2446
confusions? Cogn. Neuropsychol. 7, 161–189 evidence from semantic dementia, HSVE and a neural (2004).
(1990). network model. Brain 130, 1127–1137 (2007). 42. Mummery, C. J. et al. A voxel-based morphometry
9. Damasio, A. R. The brain binds entities and events by 26. Noppeney, U. et al. Temporal lobe lesions and study of semantic dementia: relationship between
multiregional activation from convergence zones. semantic impairment: a comparison of herpes simplex temporal lobe atrophy and semantic memory.
Neural Comput. 1, 123–132 (1989). virus encephalitis and semantic dementia. Brain 130, Ann. Neurol. 47, 36–45 (2000).
10. Damasio, A. R. & Damasio, H. in Large-Scale 1138–1147 (2007). 43. Seeley, W. W. et al. The natural history of temporal
Neuronal Theories of the Brain (ed. Koch, C.) 61–74 27. Caramazza, A. & Shelton, J. R. Domain specific variant frontotemporal dementia. Neurology 64,
(MIT Press, Cambridge, Massachusetts, USA, 1994). knowledge systems in the brain: the animate- 1384–1390 (2005).
11. Damasio, H., Grabowski, T. J., Tranel, D., Hichwa, R. D. inanimate distinction. J. Cogn. Neurosci. 10, 1–34 44. Chao, L. L., Haxby, J. V. & Martin, A. Attribute-based
& Damasio, A. R. A neural basis for lexical retrieval. (1998). neural substrates in temporal cortex for perceiving
Nature 380, 499–505 (1996). 28. Farah, M. J. & McClelland, J. L. A computational and knowing about objects. Nature Neurosci. 2,
This landmark paper combined lesion analysis in a model of semantic memory impairment: modality 913–919 (1999).
large group of patients and PET imaging in normal specificity and emergent category specificity. J. Exp. 45. Goldberg, R. F., Perfetti, C. A. & Schneider, W.
individuals to identify brain regions crucial to Psychol. Gen. 120, 339–357 (1991). Perceptual knowledge retrieval activates sensory brain
retrieval of names for entities in different semantic 29. Warrington, E. K. & Shallice, T. Category-specific regions. J. Neurosci. 26, 4917–4921 (2006).
categories (famous people, animals and tools). semantic impairment. Brain 107, 829–854 (1984). 46. Hauk, O., Johnsrude, I. & Pulvermuller, F. Somatotopic
12. Tranel, D., Damasio, H. & Damasio, A. R. A neural 30. Shallice, T. From Neuropsychology to Mental Structure representation of action words in human motor and
basis for the retrieval of conceptual knowledge. (Cambridge Univ. Press, Cambridge, UK 1998). premotor cortex. Neuron 41, 301–307 (2004).
Neuropsychologia 35, 1319–11327 (1997). 31. Jefferies, E. & Lambon Ralph, M. A. Semantic 47. James, T. W. & Gauthier, I. Auditory and action
13. Damasio, H., Tranel, D., Grabowski, T., Adolphs, R. & impairment in stroke aphasia versus semantic semantic features activate sensory-specific perceptual
Damasio, A. R. Neural systems behind word and dementia: a case-series comparison. Brain 129, brain regions. Curr. Biol. 13, 1792–1796 (2003).
concept retrieval. Cognition 92, 179–229 (2004). 2132–2147 (2006). 48. Kellenbach, M. L., Brett, M. & Patterson, K. Large,
14. Beauchamp, M. S., Argall, B. D., Bodurka, J., Duyn, 32. Jefferies, E., Patterson, K. & Lambon Ralph, M. A. colourful or noisy? Attribute- and modality-specific
J. H. & Martin, A. Unravelling multisensory integration: Deficits of knowledge vs. executive control in activations during retrieval of perceptual attribute
patchy organization within human STS multisensory semantic cognition: insights from cued naming. knowledge. Cogn. Affect. Behav. Neurosci. 1,
cortex. Nature Neurosci. 7, 1190–1192 (2004). Neuropsychologia 18 Sep 2004 207–221 (2001).
15. Snowden, J. S., Neary, D. & Mann, D. M. A. Fronto- (doi:10.1016/j.neuropsychologia.2007.09.007). 49. Kellenbach, M. L., Brett, M. & Patterson, K. Actions
Temporal Lobar Degeneration: Fronto-Temporal 33. Graham K., Patterson, K. & Hodges, J. R. Progressive speak louder than functions: the importance of
Dementia, Progressive Aphasia, Semantic Dementia pure anomia: insufficient activation of phonology by manipulability and action in tool representations.
(Churchill Livingstone, New York, 1996). meaning. Neurocase 1, 25–38 (1995). J. Cogn. Neurosci. 15, 20–46.

986 | december 2007 | volume 8 www.nature.com/reviews/neuro

© 2007 Nature Publishing Group
 REVIEWS

50. Joseph, J. E. Functional neuroimaging studies of 70. Rogers, T. T., Lambon Ralph, M. A., Hodges, J. R. & This paper demonstrates that the amodal
category specificity in object recognition: a critical Patterson, K. Natural selection: the impact of semantic semantic deficit in semantic dementia produces
review and meta-analysis. Cogn. Affect. Behav. impairment on lexical and object decision. Cogn. the same pattern of impaired performance
Neurosci. 1, 119–136. Neuropsychol. 21, 331–352 (2004). across a range of tasks differing in modality of
51. Moore, C. J. & Price, C. J. Three distinct ventral 71. Patterson, K. & Erzinçlioglu, S. in Drawing and Non- input and output.
occipitotemporal regions for reading and object Verbal Intelligence (eds Lange-Kuettner, C. & Vinter, 93. Gloor, P. The Temporal Lobe and Limbic System (OUP,
naming. Neuroimage 10, 181–192 (1999). A.) (Cambridge Univ. Press, Cambridge, in the press). New York, 1997).
52. Devlin, J. T. et al. Susceptibility-induced loss of signal: 72. Rogers, T. T. & Patterson, K. Object categorization: 94. Hickok, G. & Poeppel, D. The cortical organization of
comparing PET and fMRI on a semantic task. reversals and explanations of the basic-level advantage. speech processing. Nature Rev. Neurosci. 8, 393–402
Neuroimage 11, 589–600 (2000). J. Exp. Psychol. Gen., 136, 451–469 (2007). (2007).
These experiments highlighted an important 73. Gorno-Tempini, M. L., Cipolotti, L. & Price, C. J. 95. Adolphs, R. & Spezio, M. Role of the amygdala in
methodological issue for fMRI. Using a semantic Category differences in brain activation studies: where processing social stimuli. Prog. Brain Res. 156,
categorization paradigm and a comparison of do they come from? Proc. Biol. Sci. 267, 1253–1258 363–378 (2006).
imaging methods, the authors established that the (2000). 96. Martin, A. & Chao, L. L. Semantic memory and the
significant anterior temporal lobe activation 74. Gauthier, I., Anderson, A. W., Tarr, M. J., Skudlarski, P. brain: structure and processes. Curr. Opin. Neurobiol.
evident with H2150-PET was largely absent with & Gore, J. C. Levels of categorization in visual 11, 194–201 (2001).
fMRI — a consequence of MRI susceptibility recognition studied using functional magnetic 97. Snodgrass, J. G. & Vanderwart, M. A standardized set
artefact. resonance imaging. Curr. Biol. 7, 645–651 (1997). of 260 pictures: norms for name agreement, image
53. Mummery, C. J., Patterson, K., Hodges, J. R. & Wise, 75. Tyler, L. K. et al. Processing objects at different levels agreement, familiarity and visual complexity. J. Exp.
R. J. S. Generating ‘tiger’ as an animal name or a word of specificity. J. Cogn. Neurosci. 16, 351–362 Psychol. 6, 174–215 (1980).
beginning with T: differences in brain activation. Proc. (2004). 98. Howard, D. & Patterson, K. Pyramids and Palm Trees:
Biol. Sci. 263, 989–995 (1996). 76. Buxbaum, L. J. & Saffran, E. M. Knowledge of object A Test of Semantic Access from Pictures and Words.
54. Price, C. J., Devlin, J. T., Moore, C. J., Morton, C. & manipulation and object function: dissociations in (Thames Valley test Co., Bury St Edmunds, UK, 1992)
Laird, A. R. Meta-analyses of object naming: effect of apraxic and nonapraxic subjects. Brain Lang. 82, 99. Binder, J. R. et al. Human temporal lobe activation by
baseline. Hum. Brain Mapp. 25, 70–82 (2005). 179–199 (2002). speech and nonspeech sounds. Cereb. Cortex 10,
55. Rogers, T. T. et al. Anterior temporal cortex and 77. Chao, L. L. & Martin, A. Representation of 512–528 (2000).
semantic memory: reconciling findings from manipulable man-made objects in the dorsal stream. 100. Davis, M. H. & Johnsrude, I. S. Hierarchical processing
neuropsychology and functional imaging. Cogn. Affect. Neuroimage 12, 478–484 (2000). in spoken language comprehension. J. Neurosci. 23,
Behav. Neurosci. 6, 201–213 (2006). 78. Hodges, J. R., Bozeat, S., Lambon Ralph, M. A., 3423–3431 (2003).
56. Bright, P., Moss, H. & Tyler, L. K. Unitary vs multiple Patterson, K. & Spatt, J. The role of conceptual 101. Scott, S. K., Rosen, S., Lang, H. & Wise, R. J. Neural
semantics: PET studies of word and picture knowledge in object use: evidence from semantic correlates of intelligibility in speech investigated with
processing. Brain Lang. 89, 417–432 (2004). dementia. Brain 123, 1913–1925 (2000). noise vocoded‑speech—‑a positron emission
This publication is significant because its 79. Bozeat, S., Lambon Ralph, M. A., Patterson, K. & tomography study. J. Acoust. Soc. Am. 120,
conjunction analysis of four PET studies, using Hodges, J. R. When objects lose their meaning: what 1075–1083 (2006).
different kinds of semantic tasks and stimuli, happens to their use? Cogn. Affect. Behav. Neurosci. 102. Ferstl, E. C., Rinck, M. & von Cramon, D. Y.
revealed common activation in the anterior 2, 236–251 (2002). Emotional and temporal aspects of situation model
temporal lobe, and because it highlights the impact 80. Adlam, A. L. et al. Semantic dementia and fluent processing during text comprehension: an event-
of specificity of semantic processing. primary progressive aphasia: two sides of the same related fMRI study. J. Cogn. Neurosci. 17, 724–739
57. Crinion, J. T., Lambon-Ralph, M. A., Warburton, E. A., coin? Brain 129, 3066–80 (2006). (2005).
Howard, D. & Wise, R. J. S. Temporal lobe regions 81. Rogers, T. T., Patterson, K. & Graham, K. S. Colour 103. Noppeney, U., Price, C. J., Duncan, J. S. & Koepp,
engaged in normal speech comprehension. Brain 126, knowledge in semantic dementia: it’s not all black and M. J. Reading skills after left anterior temporal lobe
1193–1201 (2003). white. Neuropsychologia (in the press). resection: an fMRI study. Brain 128, 1377–1385
58. Scott, S. K., Blank, C. C., Rosen, S. & Wise, R. J. S. 82. Rogers, T. T. & McClelland, J. L. Semantic Cognition: A (2005).
Identification of a pathway for intelligible speech in Parallel Distributed Processing Approach (MIT press, 104. Devlin, J. T. et al. Is there an anatomical basis for
the left temporal lobe. Brain 123, 2400–2406 (2000). Cambridge, Massachusetts, USA, 2004). category-specificity? Semantic memory studies in PET
59. Lindenberg, R. & Scheef, L. Supramodal language The general theory of human semantic abilities and fMRI. Neuropsychologia 40, 54–75 (2002).
comprehension: role of the left temporal lobe for described in this book explains how we come to 105. Noppeney, U. & Price, C. J. Retrieval of visual,
listening and reading. Neuropsychologia 45, know which properties are important for auditory, and abstract semantics. Neuroimage 15,
2407–2415 (2007). representing which concepts, and why this ability 917–926 (2002).
60. Marinkovic K. et al. Spatiotemporal dynamics of may critically depend upon neuroanatomical 106. Papathanassiou, D. et al. A common language
modality-specific and supramodal word processing. convergence within the cortical semantic network. network for comprehension and production: a
Neuron 28, 487–497 (2003). 83. Plaut, D. C. Graded modality-specific specialization in contribution to the definition of language epicentres
This paper demonstrated the potential of MEG to semantics: a computational account of optic aphasia. with PET. Neuroimage 11, 347–357 (2000).
map the early temporal sequence of regional Cogn. Neuropsychol. 19, 603–639 (2002). 107. Tranel, D., Grabowski, T. J., Lyon, J. & Damasio, H.
activation in a semantic paradigm. Auditory and 84. Macario, J. F. Young children’s use of color in Naming the same entities from visual or from auditory
visual stimuli each initially engaged their respective classification: foods and canonically colored objects. stimulation engages similar regions of left
sensory cortices but ultimately the activation from Cogn. Dev. 6, 17–46 (1991). inferotemporal cortices. J. Cogn. Neurosci. 17,
each converged on the anterior temporal lobes. 85. Gelman, R. & Williams, E. M. in Handbook of Child 1293–1305 (2005).
61. Grabowski, T. J. et al. A role for left temporal pole in Psychology, Vol. II: Cognition, Perception and 108. Binder, J. R., Westbury, C. F., McKiernan, K. A.,
the retrieval of words for unique entities. Hum. Brain Development (eds Kuhn, D. & Siegler, R.) 575–530 Possing, E. T. & Medler, D. A. Distinct brain systems
Mapp. 13, 199–212 (2001). (Wiley, New York, 1998). for processing concrete and abstract concepts.
62. Gorno-Tempini, M. L. & Price, C. J. Identification of 86. Keil, F. C. Concepts, Kinds and Cognitive J. Cogn. Neurosci. 17, 905–917(2005).
famous faces and buildings: a functional neuroimaging Development (MIT Press, Cambridge, Massachusetts, 109. Mummery, C. J., Shallice, T. & Price, C. J. Dual-process
study of semantically unique items. Brain 124, USA, 1989). model in semantic priming: a functional imaging
2087–2097 (2001). 87. Murphy, G. L. & Medin, D. L. The role of theories in perspective. Neuroimage 9, 516–525 (1999).
63. Nakamura, K. et al. Functional delineation of the conceptual coherence. Psychol. Rev. 92, 289–316 110. Small, D. M., Jones-Gotman, M., Zatorre, R. J.,
human occipito-temporal areas related to face and (1985). Petrides, M. & Evans A. C. A role for the right anterior
scene processing: a PET study. Brain 123, 88. Rogers, T. T. et al. Structure and deterioration of temporal lobe in taste quality recognition. J. Neurosci.
1903–1912 (2000). semantic memory: a neuropsychological and 17, 5136–5142 (1997).
64. Gorno-Tempini, M. L. et al. The neural systems computational investigation. Psychol. Rev. 111, 111. Simons, J. S., Koutstaal, W., Prince, S., Wagner, A. &
sustaining face and proper-name processing. Brain 205–235 (2004). Schacter, D. Neural mechanisms of visual object priming:
121, 2103–2118 (1998). This paper describes a model implementation of evidence for perceptual and semantic distinctions in
65. Tsukiura, T., Mochizuki-Kawai, H. & Fujii, T. the distributed-plus-hub theory and uses the model fusiform cortex. Neuroimage 19, 613–626 (2003).
Dissociable roles of the bilateral anterior temporal to explain many different aspects of the semantic 112. Vuilleumier, P., Henson, R. N., Driver, J. & Dolan, R. J.
lobe in face-name associations: an event-related fMRI impairment observed in semantic dementia. Multiple levels of visual object constancy revealed by
study. Neuroimage 30, 617–626 (2006). 89. Garrard, P., Lambon Ralph, M. A., Hodges, J. R. & event-related fMRI of repetition priming. Nature
66. Nakamura, K., et al. Neural substrates for recognition Patterson, K. Prototypicality, distinctiveness and Neurosci. 5, 491–499 (2002).
for familiar voices: a PET study. Neuropsychologia 39, intercorrelation: analyses of the semantic attributes of 113. Phillips, J. A., Humphreys, G. W., Noppeney, U. &
1047–1054 (2001). living and nonliving items. Cogn. Neuropsychol. 18, Price, C. J. The neural substrates of action retrieval: an
67. Snowden, J. S., Thompson, J. C. & Neary D. 125–174 (2001). examination of semantic and visual routes to action.
Knowledge of famous faces and names in semantic 90. Patterson, K. The reign of typicality in semantic Vis. Cogn. 9, 662–684 (2002).
dementia. Brain 4, 860–872 (2004). memory. philos. Trans. R. Soc. Lond. B Biol. Sci. 362, 114. Brett, M. The MNI brain and the Talairach atlas. MRC
68. Hodges, J. R., Graham, N. & Patterson, K. Charting 813–821 (2007). Cognition and Brain Sciences Unit [online] http://
the progression in semantic dementia: implications for 91. Hodges, J. R., Patterson, K., Oxbury, S. & Funnell, E. imaging.mrc-cbu.cam.ac.uk/imaging/MniTalairach
the organisation of semantic memory. Memory 3, Semantic dementia: progressive fluent aphasia with (2002).
463–495 (1995). temporal lobe atrophy. Brain 115, 1783–1806 (1992).
69. Hodges, J. R., Patterson, K. & Tyler, L. K. Loss of 92. Patterson, K. et al. ‘Pre-semantic’ cognition in Acknowledgements
semantic memory: implications for the modularity of semantic dementia: six deficits in search of an We gratefully acknowledge help from E. Copper-Pye in prepa-
mind. Cogn. Neuropsychol. 11, 505–542 (1994). explanation. J. Cogn. Neurosci. 18, 169–183 (2006). ration of the manuscript.

nature reviews | neuroscience volume 8 | december 2007 | 987

© 2007 Nature Publishing Group
REVIEWS

Online correspondence
Nature Reviews Neuroscience publishes items of correspondence online. Such contributions are published at the
discretion of the Editors and can be subject to peer review. Correspondence should be no longer than 500 words
with up to 15 references and should represent a scholarly attempt to comment on a specific Review or Perspective
article that has been published in the journal. To view correspondence, please go to our homepage at: http://www.
nature.com/reviews/nrn/index.html and follow the link from the Perspective article by Stuart A. Lipton or from the
Perspective article by Kriegstein et al.
The following items of correspondence have recently been published:

Receptor abuse-dependent antagonism for neuroprotection

Hari Manev

Reply
Stuart A. Lipton

This correspondence relates to the article:

Pathologically activated therapeutics for neuroprotection

Stuart A. Lipton
Nature Reviews Neuroscience 8, 803–808 (2007)

Cause and effect in cortical folding

David C. Van Essen

Reply
Arnold R. Kriegstein, Verónica Martínez Cerdeño and Stephen C. Noctor

This correspondence relates to the article:

Patterns of neural stem and progenitor cell division may underlie evolutionary
cortical expansion
Arnold R. Kriegstein, Stephen C. Noctor and Verónica Martínez Cerdeño
Nature Rev. Neurosci. 7, 883–890 (2006)

988 | december 2007 | volume 8 www.nature.com/reviews/neuro

© 2007 Nature Publishing Group