Conceptual Structure : Helen E Moss , Lorraine K Tyler & Kirsten

words, and the activation of those features is also more consistent over different occurrences of the word. This leads to more stable patterns of activation, providing a basis for the better performance on concrete than abstract words when the input to the system is noisy, as in patients with deep dyslexia (Plaut and Shallice, 1993) or word meaning deafness (Tyler and Moss, 1997). Conversely, it is also possible that in some cases, concrete nouns may be impaired to a greater extent than abstract nouns and verbs – since the latter concepts contain fewer semantic features (especially perceptually based ones), they may be less adversely affected by damage to the semantic system (Breedin, Saffran & Costell, 1994; Saffran & Sholl, 1999). The “richer” semantic representations for concrete words also provide an account of the processing advantage for these words in the normal system when coupled with a model of word recognition which allows activation from the semantic level to feed back to facilitate processing at lower levels (e.g. McClelland and Elman, 1986, Marslen-Wilson and Welsh, 1978, Balota et al., 1991). It follows that the greater number of features activated for a given concept, the greater the amount of activation at the semantic level to facilitate semantic CONCEPTUAL STRUCTURE 10 processing, and the more semantic activation is fed back to the orthographic and phonological levels to facilitate processing here (Tyler et al., 2000b). These claims have recently been tested in greater detail in a series of studies by Pexman and colleagues (Pexman et al., 2002, 2003). For example, Pexman et al (2003) showed that semantic (concrete/abstract) decisions to words with a greater number of features (NoF) were faster than to low NoF words, a finding attributed to the greater semantic activation of high NoF words. The purported facilitatory effects of semantic feedback activation were confirmed in naming and lexical decision tasks, in which reaction times were faster to high than to low NoF words (Pexman et al., 2002), and to polysemous than nonpolysemous words (Hino and Lupker, 1996). However, semantic feedback activation is not always advantageous: words with many synonyms purportedly result in a large amount of semantic activation which feeds back to several different orthographic representations, creating competition at this level. Indeed, lexical decisions to words with many synonyms are slower than to words without synonyms (Pecher, 2001). Taken together, these elegant series of studies provide compelling evidence for a distributed, componential semantic system in which NoF plays an important role. However, it should be noted that most of the evidence on this point comes from the study of noun concepts, and similar results have yet to be demonstrated for other form classes, including verbs. In fact, one recent study suggests that (contra Fodor et al, 1980) there may indeed be a processing cost for complex event verbs over simple state verbs in reading and lexical decision times (Gennari & Poeppel, 2003). Although semantic richness or number of features is clearly an important aspect of conceptual structure, it can only be part of the story. As discussed above, concrete words typically have a greater number of features than do abstract words, affording several processing advantages in the healthy system, and providing greater feedback compensation when lexical systems are damaged. However, within the domain of concrete words, living things typically have more features than do artifacts, as shown in property generation norms (Randall et al., 2004, Garrard et al., 2001, Greer et al., 2001). Whether living things show a consistent processing advantage over non-living things in the intact system is debatable (Pilgrim et al., 2005, Laws and Neve, 1999, Gaffan and Heywood, 1993). The pattern varies with task demands; for example, in studies using pictures, living things are disadvantaged when fine-grained distinctions among similar items are needed, but this effect can be removed or even reversed when colour and texture information is added to the picture – features that are highly informative for living things (Price and Humphreys, 1989, Moss et al., 2005). Similarly, level of categorisation is important, with an advantage for living things often revealed for category level identification, but a disadvantage for basic level naming (Humphreys et al., 1988, Lloyd-Jones and Humphreys, 1997, see Moss et al., 2005 for a discussion of potential neural bases of these differences). However, for brain damaged patients, there is commonly a clear disadvantage for living things, which remains even when familiarity and other potentially confounding factors are taken into account. Thus, a high number of features alone does not protect these concepts. The basis for this discrepancy between the concrete/abstract dissociation and the living/artifact dissociation may lie in the contrasting loci of damage in these patients (both functionally and neurally). Many patients with deficits for abstract words have impairments affecting lexical systems, often limited to a specific modality of input or output (e.g. deep dyslexia, word meaning deafness). Their patterns of performance over semantic domains can be accounted for as the relative success of feedback from a largely intact semantic system to an impaired lexical system; concepts with many features providing greater support. Patients with living things deficits, on the other hand, typically have impairments to central conceptual systems, not limited to a specific modality of input or output. Although living things concepts have many features, it is CONCEPTUAL STRUCTURE 11 the nature of those features, both in terms of their distinctiveness (or lack thereof) and their correlations, that make them vulnerable to damage, as discussed in the following sections (see also Vigloccio & Vinson, this volume for a related discussion of differences in the representation of abstract and concrete words). 3.3. FEATURE DISTINCTIVENESS: FROM CUE VALIDITY TO RELEVANCE Feature distinctiveness essentially refers to the number of concepts in which a feature appears, ranging from one (e.g. has an udder, which is true of cows) to very many (e.g. has eyes, true of all animals). This factor can be traced back to the notion of cue validity – the conditional probability with which a feature signals a specific concept (Rosch and Mervis, 1975). Similarly, Devlin et al (1998) characterise features in terms of their informativeness – features occurring for very few concepts are highly informative in identifying specific concepts, while those occurring for many concepts place few constraints on the range of possible concepts to which the feature belongs. Although distinctiveness can be readily captured in distributed connectionist models of conceptual structure in terms of the overlap of activation of feature units within concept representations (e.g. Devlin et al., 1998, Durrant-Peatfield et al., 1997, Tyler et al., 2000a, Greer et al., 2001), distinctiveness has also been an important issue within the very different framework of localist hierarchical models of semantic memory. Here the drive for cognitive economy suggested that features shared by all members of a category be stored at higher levels of the hierarchy only, being “inherited” by concepts within that category, rather than duplicated for each member. Distinctive features on the other hand would need to be represented individually for each feature at lower levels (Collins and Loftus, 1975, see Moss and Marslen-Wilson, 1993 for a discussion). This was the framework in which Warrington (1975) initially interpreted the finding that patients with semantic deficits have more difficulty with distinctive than general properties, claiming that access to semantic memory may proceed in a “top down” manner starting with the general information stored at the category level, culminating in the most distinctive properties at the ends of the branches – these later, more detailed retrieval processes being more susceptible to disruption. Although the weight of evidence does not support the claim of general to specific access (Rapp and Caramazza, 1989) or the hierarchical model more generally (see introduction), the key finding that distinctive properties are more vulnerable to damage than are shared properties has proved to be a defining characteristic of semantic impairments (e.g. Bub et al., 1988, Hodges et al., 1995, Moss et al., 1997a, Moss et al., 1998, Hart and Gordon, 1992, Tippett et al., 1995). This general pattern can be accounted for within a distributed connectionist framework, given that distinctive features are experienced less frequently than are highly shared properties (which occur with many different objects) resulting in weaker connection strengths over all. However, effects of distinctiveness are unlikely to be uniform across the conceptual system. Firstly, there is considerable evidence that concepts differ in terms of the distinctiveness of their features; specifically, living things appear to have a higher proportion of shared to distinctive features than do artifact concepts. For example, Randall et al (2004) analysed the distribution of the distinctiveness of properties generated by a group of participants to 93 concepts from the categories of animals, fruit, tools and vehicles. We measured the distinctiveness of a feature as an inverse function of the number of concepts for which it is generated. Each feature has a distinctiveness value associated with it ranging from 1 (highly distinctive) to 0 (not distinctive). As predicted, the mean distinctiveness of features CONCEPTUAL STRUCTURE 12 within artifact concepts was significantly greater than that for living things. This effect has been found in a number of property generation studies, although the cut-off between distinctive and shared properties is defined in various ways across studies (Devlin et al., 1998, Garrard et al., 2001, McRae and Cree, 2002, Vinson and Vigliocco, 2002) and although the precise proportions vary across studies, largely due to the tendency for participants to underproduce highly shared features unless encouraged to do so (see Rogers et al., 2004, for a comparison across studies which highlights the potential for variation in resulting feature sets). While this domain difference in shared/distinctive feature ratio has important consequences for the effect of damage on the conceptual system, the precise consequences of the ratio of distinctive to shared properties across domains cannot be considered in isolation from other factors with which it interacts, most importantly feature type and feature correlation. Recently, Sartori & Lombardi (2004) have proposed a new feature variable of semantic relevance. Although very similar to the notion of distinctiveness, this variable is weighted according to the importance of the feature for the meaning of the concept (as derived from the number of responses in a property generation listing study and calculated using a relevance matrix). Thus, while distinctiveness is a concept-independent measure (the distinctiveness of a feature will be the same for all those concepts in which it occurs), relevance is concept-dependent (the same feature may be more relevant for one concept than another – for example, in Sartori & Lombardi’s dataset, the relevance of has a beak is greater for the concept duck than swan, largely reflecting the fact that more participants listed this feature for duck in the generation study). Sartori & Job’s analyses of the distribution of feature relevance values across categories suggest a similar pattern to that found for distinctiveness in the earlier studies; most importantly, features of living things (and especially fruit and vegetables) had significantly lower relevance values than those of nonliving things. The relevance measure may be an important development on the notion of distinctiveness as it captures the relationship between a concept and its features as a graded one, rather than in an all-or-none manner (see also Vinson and Vigliocco, 2002, for a related approach to weighting features by salience). 3.4. FEATURE CORRELATION: CLUSTERS AND MUTUAL ACTIVATION Rosch et al (1976) observed that properties of natural categories, rather than being independent, tend to cluster together, for example, that creatures with feathers generally have wings, beaks and lay eggs. Certain combinations of properties occur together much more frequently than do others. In a series of studies, Keil (1986) demonstrated that clusters of properties are larger and more densely intercorrelated for concepts within the domain of living things than of manmade objects. This notion of correlation of properties, and their variation across domains of concepts, is a central tenet of the conceptual structure approach. However, early studies of the role of property correlation in real world categories (as opposed to learning artificial concepts) were rather mixed. Malt & Smith (1984) generated a property correlation matrix for a set of basic level concepts, based on participants’ responses in feature generation and verification studies. They found that the incidence of property correlations was much greater than would be expected by chance, with about a third of all potential pairs correlated at greater than the .05 level. However, it was less clear that participants actually used this information to perform concept processing tasks: for example, adding property correlation information to a simple family resemblance weighted sum model significantly improved predictions of ratings in a typicality rating task only under certain conditions when highly salient correlations were considered and explicit comparisons CONCEPTUAL STRUCTURE 13 between correlated and uncorrelated pairs were required. This finding, along with other empirical results showing that correlations among features had little or no effects on explicit off-line category learning tasks (Murphy and Wisniewski, 1989), led to claims that conceptual processing is not typically sensitive to statistical regularities such as feature correlation, but rather that correlations tend only to be noticed when they are explicitly pointed out, or when participants are aware of a theoretical basis for why certain properties might co-occur (Murphy and Medin, 1985). This rather negative view of the role of correlation has recently been challenged by a number of studies demonstrating the importance of intercorrelation and distinctiveness in conceptual structure, both in predicting patterns of semantic impairment following brain damage and on-line activation of information in the intact system. In each case, the theoretical proposals have also been implemented in distributed connectionist models to test the validity of the major assumptions (McRae et al., 1997, McRae et al., 1999, McRae and Cree, 2002, Devlin et al., 1998, Durrant-Peatfield et al., 1997, Moss et al., 1998, Tyler et al., 2000a, Tyler and Moss, 2001). McRae & colleagues (McRae et al., 1997, McRae et al., 1999) suggest that correlations among semantic features play a role in the early computation of word meaning in on-line tasks – in contrast to the more metalinguistic conceptual reasoning tasks that had been used in earlier studies, in which higher level theoretical knowledge may be more relevant (e.g. Keil, 1989, Rips, 1989). To investigate this issue, McRae et al compared the impact of featural variables in fast on-line tasks (e.g. semantic priming, speeded feature verification) with that in slower, untimed tasks, more akin to those used in earlier studies (e.g. similarity and typicality ratings). Featural variables for concepts were established in a large-scale property generation study. Analysis of these norms supported the claim that living things have a significantly greater number of correlated properties than do artifacts (Keil, 1989), although the overall number of features did not differ across domains in this cohort. Results from the on-line semantic tasks suggested that the initial computation of word meaning is indeed highly sensitive to the distributional statistics of features within concepts. For example, in a short SOA priming task, facilitation increased as a function of the overlap in individual features for artifact pairs (i.e., number of shared features; e.g. pistol-rifle), while overlap specifically in correlated features predicted facilitation for living things (e.g. eaglehawk). However, in an untimed similarity rating task, the effect of correlation for living things disappeared, suggesting that this factor affects initial activation of the meaning rather than the eventual stable state. A similar pattern was shown in a feature verification paradigm: participants were asked to indicate whether a feature was true of a concept (e.g. deer-hunted by people). Feature correlation was manipulated such that half of the features were highly correlated with other features of the concept (e.g. hunted by people is correlated with many other properties of deer) while for the other half of the stimuli the features were presented with concepts where they were weakly correlated with other features (e.g. duck –hunted by people). Feature correlation was a significant predictor of reaction times, over and above other important factors such as conceptual familiarity and production frequency of the feature. This finding was replicated in McRae et al (1999), who also found a significant, albeit smaller effect of correlation strength at a longer SOA, and by Randall et al (2004), who reported significantly slower reaction times to the weakly correlated, distinctive features of living things than to the shared properties of living things and the relatively strongly intercorrelated properties of non-living things in a speeded feature verification task. However, correlation strength was not a significant predictor of typicality ratings in an untimed task CONCEPTUAL STRUCTURE 14 (McRae et al., 1999) nor an unspeeded feature verification task (Randall et al., 2004). Finally, McRae et al simulated the main effects of these two empirical studies in a distributed connectionist model which mapped from word form units to semantic representations which were distributed over feature units, directly reflecting the structure of concepts derived from the property generation study. In a simulation of the priming study, the model replicated the behavioural results, showing an early effect of overlap of correlated features for living things, but of individual feature similarity for artifacts. These findings suggest that one reason for null effects of correlation in the earlier category learning studies may have been due at least in part to their extended time course, which would not pick up the early effects on meaning activation that were so clearly revealed in priming and speeded feature verification tasks. 4. CONCEPTUAL STRUCTURE ACCOUNT REVISITED: CORRELATION, DISTINCTIVENESS, FEATURE TYPE AND DOMAIN Our CSA model described earlier in this chapter also stresses the combined contribution of feature correlations and distinctiveness in determining conceptual structure. A critical difference between our approach and that of Gonnerman and colleagues (Gonnerman et al., 1997, Devlin et al., 1998) and McRae and colleagues (McRae et al., 1997, Cree and McRae, 2003) is that we incorporate a set of claims about how these variables interact with differences in feature type specifically form (perceptual properties) and function in the living and non-living domains. These claims draw on the developmental literature, which investigates how children learn the relations among properties of concepts. We claim that an essential aspect of conceptual structure is the pattern of correlations between form and function (Tversky and Hemenway, 1984). If a perceptual form is consistently observed performing a function, then a system which is sensitive to co-occurrences will learn that a specific form implies a specific function (Madole et al., 1993, Mandler, 1992). The nature of these form-function relations distinguishes between living things and artifacts. Artifacts have distinctive forms, which are consistently associated with the functions for which they were created (de Renzi and Lucchelli, 1994, Keil, 1986, 1989, see also Caramazza et al.’s (1990) claim of privileged relations among properties for a similar view). Artifacts are generally designed to perform a single distinctive function so that their form is as distinctive as the function. In contrast, living things tend to ‘do’ similar things and they tend to resemble each other, thus they share many features. Individual variations in form tend not to be functionally significant (e.g. a lion’s mane). Even so, living things (like artifacts) also have form-function correlations. But whereas the form-function correlations for artifacts involve distinctive properties, for living things it is the shared properties (e.g. eyes, legs) that are involved in form-function correlations (e.g. eyes-see; legs-move). We refer to these as biological functions (Durrant-Peatfield et al., 1997, Tyler et al., 2000a, Tyler and Moss, 1997). Unlike the sensory/functional account, we do not claim that functional information is more important for artifacts than for living things, but rather that there is a difference across domains in the kind of functional information that is most strongly correlated – and therefore most robust to damage. Living things have many, very important functional properties, but the most important ones concern their biological activities that are frequently shared across most or all members of a category, rather than their intended use or purpose in relation to human beings (Tyler and Moss, 1997). These predictions were supported by the analysis of property generation norms (Randall et al., 2004). First, living things concepts had more features that were significantly more correlated with each other than were concepts in the non-living domain, a finding that has also been reported for several other property norm studies (McRae et al., 1997, Garrard et al., 2001, Vinson et al., 2003, Devlin et al., 1998). Most importantly for the CSA account, the distinctiveness of features that participated in form-function correlations (e.g., has a blade – is used for cutting) was significantly greater CONCEPTUAL STRUCTURE 15 for concepts in the non-living than the living domain. However, this finding has not been replicated in other property norm studies. Garrard et al (2001) and Vinson et al (2003) both reported a greater number of correlated features for living than non-living things, but that distinctive properties of living things were more, rather than less correlated than those of non-living things. While it is possible that both these studies were limited by the small number of concepts entered into the analyses, and by the under-estimation of shared properties overall, it will clearly be important to establish in future studies whether the interaction between distinctiveness, correlation and domain is a robust one, as claimed by the