The organisation of conceptual knowledge in the brain: The future's past and some future directions

Abstract

We review the development and current status of theories of the organisation and representation of conceptual knowledge in the human brain. The currently known facts from optic aphasia, category-specific semantic deficits, and functional neuroimaging are consistent with a framework in which the first-order constraint on the organisation of conceptual knowledge is domain. Data from functional neuroimaging suggests additionally a framework characterised by both domain- and modality-specific constraints. Work in congenital disorders and in apraxia indicate that the content of conceptual knowledge is not exhausted by modality-specific input/output processes. It is concluded that future empirical and theoretical work on the organisation and representation of conceptual knowledge will profit from a reorientation of the problem from the organisation of distinct processing systems to the content of information represented internal to such systems.

Acknowledgments

AC was supported in part by NIH grant DC04542. BZM was supported in part by a Fulbright Grant from the United States and Spanish governments. We would like to thank Erminio Capitani, Argye Hillis, Marcella Laiacona, Gabriele Miceli, Brenda Rapp, Jennifer Shelton, and Naomi Zack for many clarifying discussions. We would also like to thank Jorge Almeida, Max Coltheart, and Alex Martin for comments on an earlier version of this manuscript. We are grateful to Alex Martin for making available the graphics for Figure 3.

Notes

We use the terms “conceptual” and “semantic” interchangeably: i.e., category-specific semantic deficits (see below) could equally be called category-specific conceptual deficits. We will use the term “modality-specific semantic subsystem” to refer to information (or process) that mediates between pre- and post-semantic representations (e.g., between visual structural descriptions and phonological/orthographical lexical representations). We will refer to the latter type of representations as “modality-specific input/output” representations. The distinction between modality-specific input/output representations and modality-specific semantic representations will be examined more closely below.

Note that amodal theories of conceptual representation (e.g., OUCH) are not committed to this (unnecessarily) strong claim that the entire semantic system is involved in representing all types of information (e.g., see discussion of OUCH below).

Some early discussions of category-specific semantic deficits raised concerns that the phenomenon may be the result of uncontrolled stimulus variables: e.g., visual, complexity, familiarity, frequency (e.g., Funnell & Sheridan, 1992; Gaffan & Heywood, 1993; Stewart, Parkin, & Hunkin, 1992). Subsequent case reports in which the materials were carefully controlled for the relevant variables, as well as the observation of double dissociations over the same materials (e.g., Hillis & Caramazza, 1991) rule out this possibility (for review, see Capitani et al., 2003). More recently, Sartori and Lombardi (2004) have argued that category-specific semantic deficits for living things compared to nonliving things in a naming to definition task may arise due to a failure to “control” for the variable feature relevance, a measure of how predictable a concept is from a single feature. However, category-specific semantic deficits can manifest in a number of tasks besides naming to definition (e.g., picture naming, picture–word matching; picture matching; part–whole matching; category fluency; semantic attribute questions; see Figure 2). (For discussion and analysis of other stimulus variables, see Cree & McRae, 2003; Howard, Best, Bruce, & Gatehouse, 1995; for further discussion of the relevance framework, see Mahon & Caramazza, 2006).

One study (Gonnerman et al., 1997) reported an association between the severity of conceptual impairment and the direction of category-specific deficit, but the reported interaction has subsequently been shown to be an artifact of ranking the patients according to performance on only one object category (see Zannino et al., 2002, for discussion).

JP was impaired for knowledge of people but relatively spared for knowledge of objects and animals, while MA showed the reverse profile. Both patients were normal on a range of visuospatial tests, including object decision and perception of unfamiliar faces, and neither patient was more impaired for living or nonliving things.

This performance profile can be contrasted with that observed for case GR (Lucchelli, Muggia, & Spinnler, 1997), who presented with an anomia for people's names, but spared name recognition and face–name matching. The patient could give detailed semantic information about people he could not name, and was in the normal range for visual and verbal naming of other proper name categories (Italian cities, monuments, European cities, rivers, mountains, currencies, commercial brands).

“At minimum” because all extant theoretical interpretations assume that modality-specific input representations exist, while there is not consensus that concepts are represented independently of modality-specific input/output representations.

The Domain-Specific Hypothesis must assume that there is some innate content that allows a given domain-specific system to become “locked” to the right category of objects (i.e., a triggering mechanism). The claim is not that individual object concepts are given innately; rather, the claim is more along the lines of the type of content assumed to be localised in the theory of Martin and colleagues.

This was the case bilaterally in ventral temporal cortex, in the right superior temporal sulcus (social > mechanical), and in the left middle temporal gyrus (mechanical > social).

Such systems might be modality-context-specific as well; i.e., information is represented in a format congruent with the modality through which the information was acquired, and the only modality in which information can be stored is that through which it was acquired; see Caramazza and colleagues (1990) for discussion. The present point can be made independently of the stronger (i.e., modality-context-specific) assumption.

How the notion of “same” is fleshed out is an important aspect of simulation theory. In particular, is the simulation driven by first acknowledging the “type” of event that is occurring, and then a “type” identical simulation is run? If so, the claim would be that modality-format-specific representations are type identified. If not, then the simulation must be in terms of some particular past experience. If the former, then the question becomes how “abstract” is the information internal to modality-format-specific representations assumed to be? If the latter, then the question becomes: How does one learn anything to begin with? See the discussion of developmental data below.

For example: “Inferring intentions from observed actions might depend on the same mechanism that labels the consequences of one's own actions as being produced by one's own intentions” (Blakemore & Decety, 2001, p. 563). Or similarly, Gallese and Goldman (1998) write: “In the present article we will propose that humans' mind-reading abilities rely on the capacity to adopt a simulation routine” (p. 493).

Why were premotor areas activated only during biomechanically possible action? In other words, if activation in premotor cortex is the criterion (here) for simulation, then given that there was no such activation for biomechanically impossible actions, wouldn't this suggest that the possible/impossible classification happened somewhere else? In fact, there was some activation in orbito-prefrontal cortex that was greater for biomechanically implausible motion over plausible motion. But if the decision (or “filter”) happens outside the simulator, then the simulator is outside the system it is supposed to replace (see Blakemore & Decety, 2001, for discussion).

It might be argued that the simulation in these cases is more abstract and does not occur over representations corresponding to past experiences. At this point, however, the basic assumption of the Simulationist Framework would have been abandoned, since the claim was that information encoded during past experiences with similar events is “re-activated” in order to understand the present event.