Although the three views we have outlined are not necessarily mutually exclusive, their juxtaposition is illuminating. The empirical phenomena that motivate some perspectives may appear to challenge the underlying motivations for other points of view. Here we briefly consider how the full scope of phenomena reviewed constrains hypotheses about the broader architecture.
Explaining domain‐ and modality‐general semantic impairments First, the studies of SD that motivated the single‐hub view challenge the tripartite and many‐hubs views, which share the contention that there is no “core” conceptual representation in the cortical semantic network that contributes to processing of all kinds of concepts in all modalities of reception and expression. If knowledge is partitioned in a widely distributed network with no common core, it is difficult to understand how the relatively circumscribed pathology in SD produces such a global dissolution of knowledge.
One response to this critique is to observe that, because SD is a progressive dementia, it is difficult to be certain about the exact distribution of pathology throughout cortex. Although the structural and metabolic damage appears to be concentrated within anterior temporal regions, it may be that the pathology is broader in scope but undetectable by neuroimaging methods. On this view, the global and cross‐modal semantic impairments in SD are attributable to widespread damage throughout cortex, and not from selective damage to the ATL. There are, however, several compelling lines of evidence from neuropsychology that refute this proposal. First, other forms of frontotemporal dementia that arise from the same dis- ease process, but that mainly spare anterior temporal regions – affecting, for instance, insular or orbitofrontal cortex – produce very different cognitive impairments, with conceptual knowledge largely spared (Hodges et al., 1999; see also Chapter 20).
Second, other dementing illnesses such as Alzheimer’s disease, in which hypome- tabolism is very widespread, nevertheless do not produce comparably profound disorders of semantic knowledge (Nestor, Fryer, and Hodges, 2006; Chapter 20).
Third, recent lesion–symptom correlation studies have shown that the best predictor of impaired performance in behavioral assessments of semantic knowledge is the degree of pathology in the ATL, especially on its ventral surface (Mion et al., 2010).
Fourth, as previously noted, TMS to the ATL selectively slows semantic processing of both words and pictures, for living, nonliving, and abstract concepts. Thus the cognitive profile that uniquely characterizes SD appears to be caused by bilateral damage to the inferior aspects of the ATL, a profile that according the tripartite or many‐hubs views cannot arise from such circumscribed pathology.
Explaining modality‐specific impairments
Of course, it is the converse point – the fact that visual, verbal, and action knowledge appear to be selectively impaired or spared across different patients – that is the primary motivation for the tripartite view. If semantics depends upon a single domain‐
general cross‐modal hub, what accounts for such patterns of impairment? There are at least three responses to this question that are consistent with the single‐hub view.
Modality‐specificity arises from uncontrolled confounding factors First, patterns of apparent modality‐specific impairment can sometimes reflect confounding factors that have gone uncontrolled in the study of patients with more general semantic impair- ment. SD has been especially useful in illuminating such confounds. When knowledge degrades in this disorder, behavior across many tasks becomes increasingly influenced by modality‐specific statistical structure, and less by knowledge about semantic struc- ture. When stimuli are not matched for this statistical structure (for instance, for their orthographic or visual typicality), these factors can produce apparent dissociations across modalities. As one example, the orthographic regularity of targets and distrac- tors can strongly influence word recognition in SD: when targets in a lexical decision task are well‐formed (i.e., composed of common letter patterns) and distractors are ill‐formed, patients with severe word‐comprehension deficits can perform at ceiling discriminating words from nonwords, making it seem as though word recognition is spared when word meaning is impaired. When targets and distractors are matched for orthographic typicality, performance on the same tasks declines in proportion with word comprehension (Rogers et al., 2004b). Likewise, visual object recognition in this disorder is strongly influenced by the degree to which targets and distractors are com- posed of visually prototypical parts. In a recognition task requiring participants to dis- criminate real from chimeric animals, SD patients falsely endorsed chimeric items composed of highly prototypical animal parts (e.g., a cow with a horse’s head) and rejected real animals that possess atypical parts (e.g., a seahorse). In tests where all real objects are composed of prototypical parts and all distractors contain unusual parts, such patients can appear to have preserved recognition; but the same patients show deficits commensurate with their verbal semantic impairment when the targets and distractors are equally prototypical (Rogers et al., 2004b). Thus sensitivity to typicality or specificity may provide an alternative explanation as to why verbal, visual, and action knowledge may sometimes appear to dissociate: if typicality and specificity across tasks is not controlled, the observed dissociations may arise from such a confound.
Adlam et al. (2006) investigated the putative dissociation of lexical semantics from nonverbal semantic knowledge in patients diagnosed with fPPA – the syndrome providing perhaps the strongest support for the view that verbal semantic impairments
do not cause visual or action impairments. They examined the performance of seven newly presenting patients, all of whom met the diagnostic clinical criteria for fPPA described by Mesulam (2001), on a battery of verbal and nonverbal tasks that con- trolled or manipulated the regularity of the surface structure in the test items. Under these conditions, no patient showed the dissociation between verbal and nonverbal tasks expected under the tripartite view. In contrast to the predictions of the tripartite view, performance on verbal and nonverbal assessments of knowledge (including tests of both perceptual and action knowledge) was equally likely to be impaired; all patients showed deficits in at least some nonverbal semantic tasks; and volumetric analyses showed that both verbal and nonverbal semantic tasks were associated with the same distribution of pathology in cortex.
Modality‐specificity arises from damage to the “spokes” rather than the “hub” of the semantic network A second possibility is that, where modality‐specific patterns of impairment are not attributable to confounding factors, they arise from pathology to the modality‐specific parts of the semantic network, or to more direct mappings bet- ween these, rather than to the hub. An early demonstration of this possibility was described by Plaut (2002), who proposed a cross‐modal system of representation for encoding mappings among words, visual representations, and actions. Within this scheme, Plaut (2002) further proposed a graded anatomical division of labor, such that neurons contribute more strongly to mappings between nearby sensory‐motor representations. Thus only the neurons that are roughly equidistant from all surface representations would act as a true cross‐modal hub. Neurons closer to the “edges”
of the cross‐modal region would encode information mainly relevant to computing mappings between pairs of surface representations. Lesions to some parts of the cross‐
modal representation might selectively impair some cross‐modal mappings while leav- ing others intact, while damage to the central portion of the representation might produce a modality‐general impairment. Plaut (2002) used a computational model instantiating these principles to provide a compelling account of optic aphasia, a rare neuropsychological syndrome in which patients are unable to name visually presented objects, despite being able to (a) name them from touch and (b) retrieve their characteristic praxis from vision. This proposal represents an interesting compromise between the single‐hub and many‐hubs views.
Phonology is left‐lateralized A third hypothesis is targeted specifically at explaining why patients with SD so frequently present with profound anomia, while many of their other symptoms are less obviously apparent. In cases where pathology is very strongly left‐lateralized, anomia can be quite profound while nonverbal semantic impairment is relatively mild. Indeed, this pattern provides part of the evidence supporting the view that there exists a left‐lateralized verbal semantic system (Mesulam et al., 2003). Yet the reverse pattern, in which nonverbal semantic knowledge is seriously compromised with only mild anomia, is essentially never observed in the disorder, even in the minority of cases who present with strongly right‐lateralized pathology. Instead, such patients typically show naming and nonverbal comprehen- sion impairments that are about equal (Lambon Ralph et al., 2001).
To account for this pattern, Lambon Ralph et al. (2001) have proposed that, although the semantic hub is itself bilateral, phonological output representations are largely left‐lateralized. Thus the left hemisphere component of the hub may be more
effective at driving spoken output. In simulations with a neural network model, the authors showed that this hypothesis nicely explained the relationship between the degree of asymmetry in the neuropathology and the degree of impairment observed in naming versus other semantic task profiles.
Explaining category‐specific patterns of impairment and functional activation
A central motivation for the many‐hubs view has been the study of category‐specific semantic impairment, where patients appear to have selectively lost knowledge about a particular semantic category such as living things or manmade objects, and accom- panying brain imaging studies that appear to show selective engagement of particular cortical networks by items from different semantic categories. These empirical phenomena remain controversial, however, for at least three reasons.
First, the great majority of neuroimaging studies reporting apparent patterns of category‐specific activation have employed pictures as stimuli. Such studies are often difficult to interpret because semantic category structure is confounded with many different aspects of visual structure. Line drawings of animals are often visually more complex than line drawings of tools and other manmade objects, while tools and other manmade objects are usually much more commonly encountered in the visual environment (Funnell and Sheridan, 1992; Stewart, Parkin, and Hunkin, 1992).
Living and nonliving items may also systematically differ in the information carried across different spatial frequency bands (Coppens and Frisinger, 2005), an important factor for interpreting functional imaging data since visuospatial frequency constrains the anatomy of the visual system. Very few imaging or behavioral studies have taken care to simultaneously control for familiarity, visual complexity, and the nature of the information encoded across different spatial frequencies, and it is not clear whether apparent category‐specific patterns in these studies reflect such confounds.
Second, patterns of results in both neuropsychological and neuroimaging studies are known to be strongly influenced by the specificity or precision with which an item must be categorized in a given task. Several studies have suggested that animals from different basic‐level categories are more likely to share many properties (such as eyes, legs, the ability to move, and so on) than are manmade objects from different basic‐
level categories (e.g., cars and boats share few properties apart from the fact that they are vehicles). Thus representations of animals may be somewhat more “crowded” in general than are representations of manmade objects, so that animals are harder to differentiate from their semantic neighbors (Humphreys and Forde, 2001). Such a confound could produce both semantic impairments and patterns of functional activation that seem to be selective to animals. Consistent with this view, Lambon Ralph, Lowe and Rogers (2007) studied seven patients with herpes viral encephalitis, all of whom showed worse performance naming animals than manmade objects equated for frequency, familiarity, and visual complexity. When the same patients were asked to name animals and artifacts at a more precise level of specificity (e.g., “robin”
instead of “bird,” “ferry” instead of “boat”), the category‐specific effect disappeared:
the patients were equally impaired for both domains. The authors suggested that this occurred because, when the task requires items to be recognized very precisely, there are many close competitors in both living and nonliving domains. Similar results have
been observed in functional imaging studies of categorization at different levels of specificity (Rogers et al., 2005).
One way of bypassing such problems is to assess semantic memory with words rather than pictorial stimuli. Words do not have the same visual confounds as images, and it is much easier to control for other confounding factors such as familiarity (which can be estimated by word‐frequency counts in large corpora) and specificity (since each word is intrinsically tied to a particular level of specificity). Although a few studies employing words as stimuli have indeed yielded results consistent with the many‐hubs view (e.g., Mahon et al., 2009), such evidence appears to be relatively rare. A meta‐analysis focusing exclusively on studies that used well‐controlled word stimuli found no significant differences contrasting animals and manmade objects (Binder et al., 2009). Other meta‐analyses have failed to find consistent evidence for category‐specific activation even when including studies that employed pictorial stimuli (Joseph, 2001). Given these null results and the many potential confounding factors in the existing literature, the importance of category‐specificity for theories about the gross architecture of the semantic network remains unclear.