Abstract and concrete concepts have structurally different representat
http://www.100md.com
《大脑学杂志》
1 Dementia Research Group, Department of Clinical Neurology, Institute of Neurology, University College,2 Division of Neuroscience and Psychological Medicine, Imperial College, London, UK
Summary
The architecture supporting our conceptual knowledge of abstract words has remained almost entirely unexplored. By contrast, a vast neuropsychological, neurolinguistic and neuroimaging literature has addressed questions relating to the structure of the semantic system underpinning our knowledge of concrete items (e.g. artefacts and animals). In the context of semantic refractory access dysphasia, a series of experiments exploring and comparing abstract and concrete word comprehesion are described. We demonstrate that semantically associated abstract words reliably interfere with one another significantly more than semantically synonymous abstract words, while concrete words show the reverse pattern. We report the first evidence that abstract and concrete word meanings are based in representational systems that have qualitatively different properties. More specifically, we show that abstract concepts, but not concrete concepts, are represented in an associative neural network. Furthermore, our patient was found to have significantly greater difficulty in identifying high frequency than low frequency abstract words. This observation constitutes the first evidence of an inverse word frequency effect. Our results challenge the generality of many existing models of human conceptual knowledge, which derive their structure from experimental findings in the concrete domain alone.
Key Words: semantic refractory access dysphasia; conceptual knowledge; abstract concepts; concrete concepts
Abbreviations: RSI = response-stimulus interval
Introduction
In almost 30 years since the first publication detailing a selective impairment of semantic memory, research has revealed an ever more fine-grain categorical organization of human knowledge for concrete concepts (for a review, see Capitani et al., 2003). Over the same period, there has been a great paucity of comparable investigations of abstract conceptual knowledge. This is despite the fact that, in the original report of semantic memory deficits, the performance of two patients tested using the same stimuli revealed a double dissociation between concrete and abstract word comprehension (Warrington, 1975). The more unusual pattern of a selective impairment of concrete word knowledge with relatively preserved abstract word knowledge was subsequently replicated (e.g. Warrington and Shallice, 1984; Sirigu et al., 1991; Breedin et al., 1994). The paucity of experimental work on the abstract portion of our vocabulary has not gone unnoticed though, with Breedin et al. (1994) noting that:
‘the present results also point to a gap in the psychological literature on semantic representation, which has little to say about the manner in which abstract concepts are represented’ (p. 649).
The importance of the distinction between concrete and abstract concepts has long been recognized in a number of cognitive domains. For example, the role of item concreteness has been stressed in the context of normal adults' recall, recognition and lexicial decision abilities (Paivio, 1971; James, 1975), neurologically damaged patients' reading skills (e.g. deep dyslexics; Marshall and Newcombe, 1973), and children's vocabulary development (Brown, 1957; Gentner, 1982). Typically, subjects show a processing advantage for concrete words (e.g. leaf, house, sword) over abstract words (e.g. pity, victory, deceit). This superiority has been attributed variously to: (i) abstract words lacking the direct sensory referents of concrete words (Paivio, 1986); (ii) greater availability of contextual information in the knowledge base for concrete words (Schwanenflugel and Shoben, 1983); and (iii) concrete words being supported by more semantic features than abstract words (Plaut and Shallice, 1991, 1993). One common theme in these theories is that a quantitative distinction is drawn between concrete and abstract concepts, with concrete items more strongly represented. By contrast, we argue here that the fundamental distinction is rooted in qualitatively different principles of organization.
Theories of the representational framework underpinning conceptual knowledge have largely been restricted to the domain of concrete items, frequently being stimulated by examinations of the phenomenon of category-specific semantic deficits. Indeed, it was the observation of the selective inability of some patients to retrieve semantic information concerning certain classes of items which led Warrington (1981) to posit that semantic space had a fine-grain categorical organization. It should be noted, however, that in advocating a role for categorical information, it was specifically predicted that other types of information such as contextual association would be shown not to be powerful organizational principles:
‘I have argued that the verbal semantic systems are categorical in organization ... In putting such an argument are we rejecting any alternative model of semantic organization Yes, I think so. Such a formulation is almost certainly incompatible with notions that consider verbal semantic organization necessarily to be in the form of a network of words related by associations. To be more concrete: lion/danger or horseshoe/luck would be close in an associative network organization but distant in a categorical organization’ (Warrington, 1981; p. 422).
Nonetheless, other researchers have proposed that concepts are underpinned by an associative network in the sense that category-specific effects emerge after regional damage to an associated set of semantic features (e.g. Allport, 1985; Farah and McClelland, 1991; Hodges et al., 1994, 1995; McClelland and Rogers, 2003). Although we address the debate surrounding concrete concept organization by explicitly testing Warrington's (1981) prediction, the focus of the current experimental investigations is upon the less frequently studied knowledge of abstract concepts. In particular, we compare whether abstract and concrete word meanings are based in representational systems that have the same or different properties.
One approach to studying the properties of the semantic system underlying concrete concepts has been to examine the comprehension performance of patients with semantic refractory access dysphasia. Refractoriness is an observable, behavioural phenomenon, which can be defined by the influence of temporal factors upon performance of a cognitive task. More specifically, the performance of individuals with semantic refractory access disorders is facilitated if the delay between each response and the presentation of the subsequent stimulus is increased (response-stimulus interval; RSI).
The problem underlying refractoriness has been described as:
the reduction in the ability to utilize the system for a period of time following activationWarrington and McCarthy, 1983; p. 874.
Refractoriness has a number of behavioural consequences. First, refractoriness causes response inconsistency; with repetitive testing of a set of stimuli, identification of items may be achieved once but become unreliable thereafter. Secondly, in patients with refractory access disorders, the strong and reliable effects of word frequency are often reduced or absent. Finally, refractoriness leads to a sensitivity to the semantic relatedness of test stimuli. Refractoriness can be observed to affect not only individual concepts, but also previously unprobed concepts which are represented close in neural space. It is the last of these three consequences that has been exploited to reveal a number of very fine-grain organizational principles in the domain of concrete vocabulary (e.g. types of man-made artefacts: Warrington and Crutch, 2004) and in the domain of proper names (e.g. cities: Crutch and Warrington, 2003a).
At the physiological level, the refractory access phenomena may be caused by damage to the neuromodulatory systems which mediate neural activation levels compared with a relative sparing of the neurons which directly encode semantic information (Gotts and Plaut, 2002). This neuromodulatory damage results in excessive neuronal depression and a refractory period during which subsequent neural firing is blocked or reduced.
Refractory access disorders have also been observed to affect non-semantic components of the language system (e.g. refractory anomia: McCarthy and Kartsounis, 2000; refractory dyslexia: Crutch and Warrington, 2001). Consequently, in studying comprehension abilities in patients with refractory access conditions it is essential to verify that the refractory processes held to affect behaviour are operating at a semantic and not pre- or post-semantic level. In the case of the patient reported in this paper, A.Z., this was achieved by examining her ability to match spoken words to pictures grouped in arrays by semantic relationship (e.g. goat, deer, bear) compared with phonological relationship (i.e. rhymes such as goat, boat, coat). A.Z. made significantly more errors in the semantic than phonological condition, indicating that at least a major component of her comprehension deficit results from an impairment at the level of semantic processing (Warrington and Crutch, 2004). A semantic locus for A.Z.'s refractory access deficit is further corroborated by the presence of semantic distance effects on word–picture matching tasks when the identity of a set of items is probed under both semantically close (i.e. all potential responses drawn from the same category) and semantically distant array conditions (i.e. all potential responses drawn from different categories). Most importantly in relation to the experiments reported in this paper, equivalent semantic distance effects are also observed on tasks conducted entirely within the verbal domain, namely spoken word–written word matching. These effects provide strong evidence that the matching procedure is mediated at least in part at a semantic level (e.g. Crutch and Warrington, 2003a, 2004a).
It is this demonstration of semantic mediation on verbal matching tasks that opens the way for exploration of the domain of abstract conceptual knowledge. Indeed one reason for the dearth of abstract word experimentation is that, by their nature, low concreteness concepts have low imageability and hence cannot easily be submitted to investigation using traditional neuropsychological techniques such as picture naming or picture identification. Furthermore, the use of spoken word to written word matching tasks is unsuitable for the majority of patients with a central semantic deficit (e.g. patients with semantic dementia) because the task might still be performed without recourse to semantic mediation via a process of print–sound correspondence. As noted though, these techniques can and have been used to explore fine-grain categories within a wide-ranging vocabulary in patients with a refractory access disorder of semantic processing. Here the technique is applied for the first time to abstract concepts.
a more practical perspective, semantic distance effects have typically been measured in the concrete domain by comparing patients' abilities to identify target items from among distractor items drawn from the same semantic category (e.g. celery, radish, leek, onion) as opposed to distractors drawn from different semantic categories (e.g. celery, rhino, flask, helicopter). In the case of patients with a semantic refractory access deficit, refractoriness is observed to build up much more quickly in the former than in the latter condition (e.g. Forde and Humphreys, 1995; Warrington and Cipolotti, 1996; Warrington and Crutch, 2004). However, most abstract words are not as easily classified as concrete words. For example, the word ‘dog’ can be classed as an animal (superordinate) and has various exemplars such as ‘labrador’ and ‘alsatian’ (subordinates), but the word ‘luck’ cannot easily be categorized in the same manner. Many abstract words do, however, have synonyms, i.e. other abstract words which have essentially the same meaning (e.g. reluctant, unwilling, averse). Indeed, in the studies reported here we use synonymy as a marker of semantic similarity among abstract words.
In the context of a semantic refractory access disorder, this series of experiments attempts to measure the relative decrement in comprehension accuracy as influenced by similarity and association of meaning. Thereby, we attempt to compare directly semantic similarity and semantic association as competing principles of organization of both abstract and concrete word semantics. In doing so, we document the effects of abstractness and frequency upon the comprehension abilities of our patient.
Case report
A.Z. is a 73-year-old right handed lady who became globally aphasic following a left middle cerebral artery haemorrhage in January 2002. A CT scan demonstrated extensive damage to the left middle cerebral territory with left parietal, temporal and posterior frontal involvement. A.Z.'s spontaneous speech was fluent yet jargonistic with frequent semantic and phonemic errors. She also demonstrated a gross and pervasive impairment of her reading, writing and naming skills. In particular, she was unable to read high frequency monosyllabic words, could not write her own name and named only two items from the Boston Naming Test.
With regard to her comprehension skills, A.Z. exhibited a mild static semantic impairment. On the short form of the British Picture Vocabulary Scale (Dunn et al., 1982), she scored 22 out of 32 on oral presentation and 23 out of 32 with written stimuli. Furthermore, on the Pyramid and Palm Trees Test (Howard and Patterson, 1992), A.Z. gained the weak score of 38 out of 52 on both the word–word matching and picture–picture matching versions of the task. Comprehension performance on a high frequency word-picture matching task revealed no evidence of a category-specific semantic deficit (McKenna and Warrington, 1978).
Therefore, despite having no viable expressive language skills, systematic investigation of A.Z.'s comprehension abilities remained possible with the use of simple written word or picture pointing tasks. These techniques have previously been applied to three aspects of her conceptual knowledge, namely place names (Crutch and Warrington, 2003a), person names (Crutch and Warrington, 2004) and modality-specificity (Warrington and Crutch, 2004). In this paper, however, we tackle a much larger and arguably more important component of conceptual knowledge, namely the realm of abstract words. The investigations reported here were conducted between June 2003 and January 2004, and were designed to document and compare A.Z.'s comprehension of abstract and concrete vocabularies.
Experiment 1: Frequency and abstractness effects in abstract word identification
Words are often classified as ‘abstract’ or ‘concrete’, particularly in the context of neuropsychological experimentation. However, abstractness/concreteness may in fact constitute a continuum ranging between highly abstract concepts (e.g. tendency) and highly concrete concepts (e.g. spoon). If this were the case, then a number of reports of concreteness effects in patients may reflect a greater expression or an exacerbation of a pre-existing abstractness gradient. In this experiment, we explored the influence of the degree of abstractness and word frequency upon A.Z.'s word identification skills. More specifically, abstractness was manipulated without resorting to stimuli drawn from the concrete extreme of any potential continuum; only words traditionally classified as abstract were included in the experiment.
Methods
The stimuli consisted of two sets of 144 abstract written words with frequency ratings and abstractness ratings drawn from the MRC Psycholinguistic Database (Coltheart, 1981; see Appendix 1). All words were between five and eight letters in length. The abstract words in both sets were divided equally into a matrix of three levels of frequency (low, middle and high frequency) x three levels of abstractness (highly abstract, moderately abstract and mildly abstract). For the frequency ratings, there was no overlap between the matrix cells, with each frequency band defined by a strict range of Thorndike–Lorge ratings (low: 1–15; middle: 100–200; high: >300). Similarly for the abstractness ratings, there was no overlap between the ratings that defined each band (very low: 200–300; low: 300–400; mild: 400–500). The 16 items in each matrix cell of each set were arranged into four written word arrays of four items. The experiment was conducted as a spoken word–written word matching task, so that for each group of four arrays, one item was probed from each array in sequence until every item had been probed once. A.Z. was required to point to the target written word following each spoken word stimulus. The experiment was conducted at a natural presentation rate with a stimulus-response interval of 1 s. For the first set of words, low, middle and high frequency arrays were tested in order for the very low abstractness words, and latterly for the low and mildly low words. For the second set of words, the identical procedure was used except that arrays corresponding to each cell of the matrix were presented in the reverse order. More specifically, the high frequency/mild abstractness cell was tested first and the low frequency/very low abstractness cell was tested last.
Results
The percentage correct responses for each cell of the word matrix (both stimulus sets combined) together with the percentage total correct responses for each frequency and abstractness band are shown in Table 1. A 2 test for trend of odds revealed that A.Z. showed significant trends to better performance not only with less abstract words [2(1) = 8.95, P < 0.01] but also with lower frequency words [2(1) = 4.56, P < 0.05].
Comment
These results demonstrate that A.Z. shows an increase in impairment in written word identification the greater the abstractness of the target word. However, the pattern of performance is atypical in the sense that the results reported here document for the first time the reversal of the typical word frequency effect. Patients with semantic refractory access disorders have previously been shown to be relatively insensitive to the effects of word frequency (i.e. the absence of a frequency effect), but have never before exhibited an actual reversal of the tendency. The superiority of low frequency word performance is particularly striking, as an advantage for high frequency words is almost ubiquitous in clinical and experimental studies of all components of the language processing system.
Experiment 2: The effect of temporal factors upon abstract word comprehension
In a number of previous word comprehension studies, we have demonstrated that A.Z. exhibits the core features of a refractory access deficit. In particular, A.Z. shows a sensitivity to temporal factors, with performance being facilitated by increasing the length of the interval between making each response and the presentation of the subsequent stimulus (RSI). However, these demonstrations have been conducted in the domains of concrete concepts and proper nouns. Therefore, in order to study semantic distance effects in A.Z.'s comprehension of abstract words, it was considered necessary to establish whether the influence of temporal factors was also apparent in this domain of conceptual knowledge.
Methods
The stimuli consisted of 20 written abstract words, arranged into five arrays each containing four items (e.g. wait, task, ache, stiff; see Appendix 2). For each array, the name of every written word was spoken by the examiner four times in a pseudo-random order. Following the presentation of each spoken word target, A.Z. was requested to point to the corresponding written word in the array. This procedure was conducted twice for each array, once under a fast presentation rate (1 s RSI) and once under a slow presentation rate (10 s RSI). The two temporal conditions were alternated in an ABABA BABAB design.
Results
The results for each condition are shown in Fig. 1. A Wilcoxon matched-pairs signed-ranks test showed that A.Z. made significantly more errors in the fast condition than in the slow condition (z = 2.50, P < 0.02; Fig. 1A). Having probed each item four times, an item-specific analysis was conducted to examine on which probe each error occurred. Serial position curves showing the percentage error rate on each probe are given in Fig. 1B. In the fast condition, a Cochran Q test revealed a significant difference in the number of errors made on each probe [Q(3) = 16.80, P < 0.001]. This difference reflects a gradual increase in error rate with successive probes following initially accurate performance. In the slow condition however, A.Z.'s performance was almost at ceiling and there was no significant difference in the number of errors on each probe [Q(3) = 3.62, P > 0.3]. Thus, A.Z.'s ability to comprehend the meaning of abstract words is characterized by at least two of the core features of a refractory access disorder, namely sensitivity to temporal factors and inconsistency.
Experiment 3: The influence of semantic similarity upon abstract adjective and verb comprehension
The exploration of semantic organization in the context of refractory access dysphasia has typically depended upon the manipulation of semantic distance effects. In particular, it has been shown that, despite accurate performance initially, refractoriness builds up more quickly between semantically related concepts than semantically unrelated concepts leading to a higher overall error rate. Semantic similarity has been demonstrated to affect concrete word comprehension performance not only in patient A.Z., but also in other semantic refractory access patients (e.g. Forde and Humphreys, 1995; Warrington and Cipolotti, 1996). However, equivalent effects of semantic similarity in the abstract word domain have not previously been investigated. In this experiment, comprehension of two classes of abstract words was investigated, namely adjectives and verbs. Unlike in the concrete domain in which ‘semantic relationship’ is most frequently defined by membership of a common semantic category, few abstract concepts can be classified in the same manner. Consequently, we examined semantic distance among adjectives using synonyms derived from a thesaurus to form semantically similar and semantically dissimilar groups. By contrast, the verbs selected in this task were not synonymous but were chosen because they formed coherent, semantically related clusters of concepts.
Methods
The stimuli consisted of two sets of 16 adjectives and one set of 16 verbs (see Appendix 3). The items within each adjective set were arranged into four semantically related written word arrays (e.g. divine, sacred, religious, holy). The same words were also arranged into four semantically unrelated arrays, each containing one word from each the semantically related arrays (e.g. divine, sparse, furious, sudden). In an identical manner, the items within the verb set were also arranged into both semantically related (e.g. kneel, sit, stand, lie) and semantically unrelated written word arrays (e.g. kneel, sing, bake, drive). The tasks were conducted as spoken word–written word matching tests. Each item in an array was probed four times in a pseudo-random order and a 1 s RSI was adopted. Comprehension of each set of words was tested separately and, within each set, the arrays were presented in an ABBA order.
Results
The percentage correct responses for adjectives (both word sets combined) and verbs under the conditions of semantic relatedness and unrelatedness are given in Fig. 2A. Despite the semantic similarity of words in the related arrays, Wilcoxon matched-pairs ranked-sums tests revealed no difference between A.Z.'s ability to identify adjectives or verbs under the two conditions (adjectives: z = 1.16, P > 0.2; verbs: z = 0.93, P > 0.3). Serial position curves demonstrating the percentage error rate for each probe of adjective and verb items are shown in Fig. 2B. Cochran Q tests revealed that A.Z. showed a significant difference in the number of errors made on successive probes with both the adjectives [related: Q(3) = 29.40, P < 0.001; unrelated: Q(3) = 17.89, P < 0.001] and the verbs [related: Q(3) = 8.48, P < 0.05; unrelated: Q(3) = 10.00, P < 0.02]. Examination of Fig. 2B indicates that these differences reflect increasing error rates with repeated probes following an almost ceiling level of performance on probe 1.
Comment
Overall, the data reported in this experiment suggest that refractoriness does not build up more quickly among semantically related abstract words more than semantically unrelated words. The equivalence of performance on related and unrelated arrays of both adjectives and verbs was unexpected, and very different to the pattern of performance predicted for concrete nouns.
Experiment 4: The influence of semantic similarity upon abstract and concrete word comprehension
In the context of a semantic refractory access dysphasia, semantic distance effects are attributed to the abnormal, deleterious effects which activating a concept has upon other concepts that partly share neural space. Thus, the failure to observe semantic distance effects among abstract words in Experiment 3 was highly surprising, not only because such effects are almost ubiquitous within the concrete domain, but because the observation implies that abstract words with similar meanings do not necessarily share neural space. In this experiment, the influence of semantic similarity upon A.Z.'s comprehension abilities was further investigated by comparing and contrasting abstract and concrete words. Given that there are very few adjectives with high concreteness values, the abstract words chosen for this experiment were mainly nouns and verbs.
Methods
The stimuli consisted of two sets of abstract words (see Appendix 4) and two sets of concrete words (see Appendix 5). Within each set of abstract words, the items were arranged into semantically synonymous written word arrays each containing four abstract synonyms (e.g. deceit, trick, steal, cheat). The same words were also arranged in semantically non-synonymous arrays consisting of one word from each of four synonymous arrays (e.g. deceit, strike, mush, screen). In an identical fashion, the items within each set of concrete words were arranged into four written word arrays in which each item was drawn from the same semantic category (e.g. goose, pigeon, crow, sparrow). The same concrete words were also arranged into four different semantic category arrays containing one item from each of the same category arrays (e.g. goose, melon, pullover, biscuit). All other aspects of the experimental design were identical to the procedure employed in Experiment 3.
Results
The combined percentage correct responses for the synonymous and non-synonymous abstract word conditions are shown in Fig. 3A. A Wilcoxon matched-pairs signed-ranks test revealed no difference in response accuracy between the synonymous and non-synonymous array conditions (z = 0.22, P > 0.8). Serial position curves demonstrating the percentage error rate for each probe of the abstract words are also shown in Fig. 3A. Cochran Q tests revealed that A.Z. showed a significant difference in the number of errors made on successive probes [synonyms: Q(3) = 17.48, P < 0.001; non-synonyms: Q(3) = 12.00, P < 0.01]. Inspection of the serial position curves suggests that these differences reflect an increase in the percentage error rate across consecutive probes of the test stimuli for both synonymous and non-synonymous conditions alike.
The percentage correct responses on the concrete word stimulus sets are shown in Fig. 3B. The results for both word sets were considered together. A Wilcoxon matched-pairs signed-ranks test showed that A.Z. was significantly less accurate at pointing to the target word in same semantic category arrays than different semantic category arrays (z = 3.62, P < 0.001). Serial position curves showing the percentage error rate for successive probes of the test items under both same category and different category conditions are also given in Fig. 3B. Cochran Q tests again demonstrated a significant difference in response accuracy between probes under the same category condition [Q(3) = 14.76, P < 0.01] and a trend towards a similar pattern in the different category condition [Q(3) = 7.43, P < 0.1]. Both these results describe a gradual increase in the error rate across successive probes.
Comment
The evidence presented in this experiment strongly suggests that refractoriness does not build up more quickly among abstract words which have similar meanings than among those which do not. By contrast, significant effects of semantic relatedness were observed in the comprehension of more concrete words. The surprising absence of the semantic relatedness effect among abstract words cannot be attributed to the grammatical class of the abstract word stimuli in question as equivalent null results were obtained using adjectival stimuli as when using noun and verb stimuli.
Experiment 5: The influence of semantic association upon abstract and concrete word comprehension
Having established that abstract word representations are unaffected by semantic similarity (i.e. category) relative to concrete words, we evaluated an alternative potential principle of organization: semantic association. Associated words are those whose meanings are not synonymous but which are often bound together in real world or sentential contexts (e.g. salute, army, general, respect). The significance of associative links between concepts was examined using the same spoken word to written word matching technique described in Experiments 3 and 4.
Methods
The stimuli consisted of two sets of 16 abstract words (see Appendix 6) and two sets of 16 concrete words (see Appendix 7). For all stimulus sets, the items were arranged into arrays each containing four associated but non-synonymous words (e.g. abstract: exercise, healthy, fitness, jogging; concrete: farm, cow, tractor, barn). The same words were also arranged into four non-associated arrays with one word drawn from each associated array (e.g. abstract: exercise, gamble, punch, future; concrete: farm, sailor, shelf, oven). As before, the tasks were conducted as spoken word–written word matching tests. Each item in an array was probed four times in a randomized sequence, using a 1 s RSI. Comprehension of each set of words was tested separately and, within each set, the arrays were administered in an ABBA design.
Results
The percentage correct responses (sets 1 and 2 combined) for the abstract words under both associated and non-associated array conditions are shown in Fig. 3C. A Wilcoxon matched-pairs signed-ranks test showed that A.Z. showed a difference between conditions with significantly worse performance on the associated word arrays than non-associated word arrays (z = 4.23, P < 0.001). Serial position curves showing the percentage error rate across the four probes of each test item are also given in Fig. 3C. Cochran Q tests demonstrated a strong trend towards different accuracy levels between probes in the associated array condition [Q(3) = 6.48, P < 0.1] and a significant difference between probes in the non-associated condition [Q(3) = 15.67, P < 0.01]. Inspection of the serial position curves suggests that both these effects correspond to an increase in the percentage error rate across consecutive probes of the test stimuli.
The percentage correct responses on the concrete word stimulus sets are shown in Fig. 3D. The results for both word sets were considered together. A Wilcoxon matched-pairs signed-ranks test revealed no significant difference between the associated and non-associated conditions (z = 0.23, P > 0.8). As with the abstract word stimuli, serial position curves showing the percentage error rate for successive probes of the test items are also given in Fig. 3D. Cochran Q tests also demonstrated significant differences in response accuracy between probes under both the associated array condition [Q(3) = 15.42, P < 0.01] and the non-associated array condition [Q(3) = 22.83, P < 0.001]. Both these results describe a gradual increase in the error rate across successive probes.
Comment
The significant effect of semantic association upon the comprehension of abstract words makes a striking contrast to the absence of semantic similarity effects in the abstract synonyms tasks. It appears that refractoriness builds up much more easily among associated than synonymous abstract words, implying that the representations underpinning associated abstract words are much closer neural neighbours than the representations which support words which have very similar meanings. In the concrete domain, evidence of the reverse pattern has been documented. The absence of an effect of semantic association upon the comprehension of concrete words stands in contrast to the strong effect of semantic similarity. Thus these results give further strength to the claim that, unlike abstract concepts, concrete concept representations are organized by semantic category and not by semantic association.
Discussion
The observation and manipulation of semantic relatedness effects in patients with semantic refractory access disorders enable the investigation of principles governing the organization of semantic memory. It is held that refractoriness may be observed between related concepts because the activation of one representation leads to the partial activation of other representations that share neural space. In previous reports of our patient A.Z., we have described a number of such effects in the conceptual domains of concrete items and proper nouns (Crutch and Warrington, 2003a, 2004, 2005; Warrington and Crutch, 2004). In this paper, we have extended this approach to the rarely studied domain of abstract words. Sets of synonymous abstract words were selected to be analogous to the semantically related categorical groupings of concrete items employed in earlier studies. However, to our great surprise, we were entirely unable to detect semantic relatedness effects among synonymous abstract word stimuli. This unexpected absence of a relatedness effect was documented for several sets of stimuli. It was demonstrated that adjectives (Experiment 3), verbs (Experiment 3) and also mixed noun and verb parts of speech arrays (Experiment 4) were all equally immune to the effect. By contrast, using an identical spoken word–written word matching procedure, strong semantic distance effects continued to be observed among concrete words (Experiment 4). The unpredicted null results led us to consider alternative forms of relationship between abstract words and, in particular, contextual associative relationships. On a task comparing A.Z.'s ability to identify associated and non-associated abstract words, A.Z. showed significantly worse performance in the associated condition (Experiment 5). However, remarkably no effect of semantic association was detected on an equivalent task involving concrete words (Experiment 5). Thus, we have documented evidence of a double dissociation between abstract words which show an effect of semantic contextual association but not semantic similarity, and concrete words which show an effect of semantic similarity but not semantic association. On the basis of these contrasting patterns of results, we would claim that abstract concepts are represented in an associative neural network whereas concrete concepts have a categorical organization. Indeed, while some researchers have claimed that concrete concepts are imbedded in a hierarchy of superordinate and subordinate representations, this principled organization appears not to apply to a more abstract vocabulary. We would predict that these contrasting organizations must be reflected in the neural substrate of conceptual knowledge.
In this discussion, we first consider existing views in the cognitive neuroscience literature upon the difference between abstract and concrete concepts before considering the implications of our own results. In both of these considerations, we make a clear distinction between theories of how the abstract–concrete difference is acquired and theories of how that difference is represented.
It has been claimed that the abstract–concrete difference reflects the manner in which these concepts are acquired and the roles they play in language (Breedin et al., 1994). Experience of the perceptual features of objects via our five sensory channels appears to play a key role in the acquisition of concrete concepts (e.g. Landau et al., 1992, 1998). Abstract concepts, however, may be acquired in the context of language without any direct perceptual input. As a consequence, Breedin et al. (1994) suggested that: ‘exposure to multiple sentence contexts plays an important role in the development of representations for abstract concepts’ (p. 650).
Consistent with this viewpoint, abstract words have been claimed to have richer associations, and indeed more meanings, than concrete words (Paivio et al., 1968; Schwanenflugel and Shoben, 1983). In particular, the meaning of concrete words (e.g. horse) tends to be constant across most contexts, whereas abstract words (e.g. appeal) often have related but distinct meanings depending on the context in which they are being used. Experimental support from this notion comes from Schwanenflugel and colleagues who in a series of timed comprehension, naming and meaningfulness judgement tasks showed that the provision of sentence or paragraph contexts eliminated the traditional bias toward faster responses with concrete items (Schwanenflugel and Shoben, 1983; Schwanenflugel et al., 1988, Schwanenflugel and Stowe, 1989).
The question of whether abstract words are represented in a qualitatively or quantitatively different way from concrete words has been considered as an explanation of some abstractness effects (e.g. abstract word anomia; Franklin et al., 1995). However, theoretical accounts of conceptual organization have tended to either concentrate upon quantitative differences between the two types of words or to only model concrete concepts to the exclusion of abstract knowledge. For example, Plaut and Shallice (1993) hypothesized that the difference was a reflection of concrete words being supported by a greater number of semantic features. To examine this notion, they developed a connectionist model of oral reading via semantics in which lesions at any locus produced the typical concreteness effect (i.e. more errors for abstract words). However, this result would appear to be unsurprising given that an a priori decision was taken to train the model on a greater number of semantic features for concrete than abstract terms. By contrast, the model of Allport (1985) only considers meaning for imageable words, positing the existence of attribute domains which correspond to different sensory properties. For those concrete concepts for which the model does offer a system of organization, the underlying representational framework is held to be associative in nature, with connections between distributed features according to their co-occurrence.
To our knowledge, only one previous report has claimed, as we do, that the representational difference between abstract and concrete words might be qualitative rather than quantitative (Breedin et al., 1994). However, these researchers characterized the difference in terms of discrepancy in the number of perceptual and functional features required to distinguish abstract and concrete concepts. Our results suggest that irrespective of the type of information which constitutes abstract and concrete concepts, a much more fundamental difference exists in the actual architecture of those representations. The evidence that the spread of refractoriness between concrete concepts is exacerbated by semantic similarity but not semantic association confirms Warrington's prediction that in the domain of concrete concepts, the primary organizational principle is categorical and not associative (Warrington, 1981). By contrast, it appears that the representations of abstract concepts are predominantly linked to other associated rather than semantically similar items. This conclusion is compatible with a wide range of previous observations about the effects of word concreteness, particularly those of Schwanenflugel and colleagues detailed above. We would regard the facilitatory effect of sentential contexts as a further example of the importance of the associative links between abstract concepts. However, we should be careful to note that, given the probable existence of a continuum between abstractness and concreteness, we suggest this associative/categorical dichotomy is relative rather than absolute. More specifically, we predict that middle concreteness items (e.g. nurse, chemistry) have both associative and categorical connections, in more equal proportions than concepts at either end of the spectrum. It is difficult to envisage how existing models of conceptual knowledge, which have mainly been derived from work on category-specific deficits of concrete item knowledge, can be extended to incorporate our findings on abstract words. Essentially, our findings suggest that attempting to model conceptual knowledge within a unitary system based on a single set of network principles is over simplistic.
With regard to the development of different representational frameworks for concrete and abstract concepts, we agree with Breedin et al. (1994) that multiple contexts may play a major role. It is notable that in a dictionary, concrete items typically have only a few fixed meanings (e.g. tomato occurs in the context of gardening and eating). On the other hand, dictionaries tend to describe abstract words as having multiple meanings depending upon context (e.g. the word ‘set’ has >30 transitive and intransitive senses). Furthermore, the psycholinguistic literature makes a clear distinction between two types of ‘multiple meaning’, namely homonyms which have clearly distinct, unrelated meanings (e.g. ‘clog’ = Dutch footwear, to block up, etc.) and polysemous words which have multiple, semantically related senses (e.g. ‘pain’ = punishment, grief, to cause distress, etc.). Recent evidence from lexical decision tasks in normal subjects has demonstrated the importance of this latter distinction by showing a processing disadvantage for words which have multiple unrelated meanings, but a processing advantage for words which have multiple related senses (Rodd et al., 2002, 2004). However, it should be noted that some researchers have suggested that all words are polysemous to some extent, because their precise meanings change in every unique context in which they appear (Anderson and Nagy, 1991). We would argue that an associative network may have a greater capacity to represent the multiple meanings of abstract words than a more formal, rigid categorical network. Such a representational framework also possesses the flexibility to capture the natural variability in the semantics of both words with unrelated meanings and words with related senses.
One further finding which is of particular note is that our patient A.Z. not only exhibited a commonly observed concreteness effect on written word identification tasks, but also showed some evidence of an inverse frequency effect (Experiment 1). An absence of an advantage for high frequency words has been previously reported in several studies of both refractory access patients (e.g. Warrington and Cipolotti, 1996) and patients with static storage impairments of various components of the language system (e.g. Hillis et al., 1990; Nickels and Howard, 1995). Indeed, in another study of our patient A.Z., frequency effects were noted to be absent on a spoken word–picture matching task (Crutch and Warrington, 2004b). However, to the best of our knowledge, A.Z. is the first patient on record to demonstrate significantly better performance with less commonly occurring stimuli. One previous patient exhibiting a pre-disposition for low frequency word production has been reported, but this was in the context of a classical jargon aphasia in which the nature and information processing locus of the impairment was unclear (Marshall et al., 2001). The absence of frequency effects in refractory access patients has been attributed to an impairment of the neuromodulatory system leading to higher than normal levels of synaptic depression following the activation of a representation. Frequency effects are reduced or absent because synaptic depression is greater when activity levels are higher, as in the case of high frequency items (Gotts and Plaut, 2002). We would claim that this explanation can be extended to an actual inversion of the frequency effect; using very low and very high frequency stimuli as in Experiment 1 may generate a sufficiently large activity discrepancy as to generate significantly more synaptic depression following high frequency than low frequency word identification. It should again be stressed that the inversion of the frequency effect in A.Z. is not necessarily reliable and is very much stimulus-dependent. On this account, an absence rather than an inversion of the frequency effect was observed in our earlier study of the same patient (Crutch and Warrington, 2004b) because the (pictureable) stimuli invoked a smaller frequency range.
To summarize, we have reported a series of experiments which indicate that abstract conceptual knowledge relies upon qualitatively different representations than concrete conceptual knowledge. We would suggest that a deeper understanding of abstract word semantics and their relationship to other aspects of the language system is long overdue, particularly bearing in mind the importance of an abstract vocabulary to our ability to communicate and produce fluent propositional speech (Crutch and Warrington, 2003b). Overall, studies of concrete word representations have yielded valuable insights into the structure of human semantic memory, but it is important to bear in mind that they concentrate upon only a fraction of our total vocabulary.
Appendix 1 Stimuli contained in both Set A and Set B of the abstractnessxfrequency matrix task (Experiment 1)
Very low abstractness words (concreteness rating 200–300)
Low frequency Middle frequency High frequency
Set A Set B Set A Set B Set A Set B
avowal dotage concert approval advice respect
excise impunity motive misery emotion delight
truism throe boast honesty blame attitude
feint pacifism tendency anxiety ideal value
savant infinity loyalty devotion pride variety
levity clemency virtue economy quality trust
enigma sagacity distress sorrow instance opinion
alacrity jeopardy wisdom patience memory spite
analogy atrocity mercy extreme purpose success
overtone reprisal magic despair truth spirit
inaction sanctity upset estimate effort manner
heroism protocol talent nonsense guess effect
alias craven dread capacity feeling cause
unbelief betrayal impulse reality chance least
impetus heresy neglect extent while might
bravery aptitude dignity mystery other thought
Low abstractness words (concreteness rating 300–400)
Low frequency Middle frequency High frequency
Set A Set B Set A Set B Set A Set B
emission stipend culture insult profit claim
calculus pommel retreat skill charm distance
prefix pronoun humanity rescue fortune century
tenure perjury welfare threat strain occasion
stoic sequel flutter session fashion protest
regency axiom crisis amateur strength pleasure
sobriety legality victory transfer relief peace
jargon truce portion menace amount health
adultery carnage incline comedy beauty position
bequest rigidity bargain grief result marriage
context velocity proof angel number dream
acidity edict poverty division problem month
interim deity drama research reason sense
foible mirage budget liberty answer trouble
dogma equity incident reaction moment minute
mutiny agility genius reward right business
Mild abstractness words (concreteness rating 400–500)
Low frequency Middle frequency High frequency
Set A Set B Set A Set B Set A Set B
buffoon oxide shriek footstep speech leader
zenith curfew comrade craft patient degree
vitriol molecule fairy ceremony murder object
pestle abbess embrace grate income accident
galaxy mermaid saint rebel shadow youth
abyss fissure stalk deposit breath design
dynasty glutton adult glitter whisper meeting
venom ignition pioneer resort report travel
cuisine psalm scent autumn summer general
mileage donor device disaster notice enemy
alphabet citation signal lecture figure group
segment mosque groan exhaust party fight
shale maize angle damage matter story
quake rogue route bunch laugh country
brawl microbe holiday shrug evening voice
filth sedative interior payment night friend
Appendix 2 Stimulus arrays tested under both fast and slow presentation conditions (Experiment 2)
rest ache tough work great
job tall steep pain task
firm chore wait high hard
sore halt hurt stiff stop
Appendix 3 Adjective and verb stimulus arrays used in Experiment 3
Adjectives (Set 1)
Synonymous arrays
divine scant enraged sudden
sacred bare furious abrupt
religious sparse annoyed rash
holy deficient angry unexpected
Non-synonymous arrays
scant furious sparse angry
sacred rash holy sudden
annoyed deficient enraged bare
unexpected divine abrupt religious
Adjectives (Set 2)
Synonymous arrays
hurried silent stupid glad
fast quiet dense happy
quick peaceful thick cheerful
speedy still foolish merry
Non-synonymous arrays
merry silent dense speedy
foolish stupid quiet thick
quick glad happy cheerful
still hurried fast peaceful
Verbs
Semantically related arrays
sing kneel bake drive
shout sit cook fly
laugh stand boil sail
talk lie fry ride
Semantically unrelated arrays
sing shout stand ride
kneel sit fry talk
bake cook laugh boil
drive fly sail lie
Appendix 4 Stimulus arrays probed in synonymous and non-synonymous conditions of the abstract synonym task (Experiment 4)
Set 1
Synonymous arrays
boil look gale clean eat
heat peek wind neat chew
cook glance storm pure bite
fry see breeze fresh taste
Non-synonymous arrays
boil heat neat cook see
look peek chew storm taste
gale wind fry pure fresh
clean eat glance bite breeze
Set 2
Synonymous arrays
beat mask cheat soft
knock cloak trick mush
thump hide steal pulp
strike screen deceit smooth
Non-synonymous arrays
beat smooth thump screen
cloak mask hide strike
cheat trick steal deceit
soft knock pulp mush
Appendix 5 Stimulus arrays probed in same category and different category conditions of the two word sets in the concrete categories task (Experiment 4)
Set 1
Same category arrays
goose cardigan pizza melon
crow jacket cake grapes
sparrow blouse bread peach
pigeon pullover biscuit banana
Different category arrays
goose banana pigeon cake
melon cardigan bread blouse
pullover pizza jacket crow
biscuit sparrow grapes peach
Set 2
Same category arrays
kettle rabbit leek daffodil
toaster hedgehog carrot tulip
oven squirrel onion rose
fridge badger parsnip pansy
Different category arrays
squirrel badger rabbit tulip
onion rose kettle oven
toaster parsnip daffodil carrot
pansy fridge leek hedgehog
Appendix 6 Stimulus arrays probed in associated and non-associated conditions of the two word sets in the abstract associates task (Experiment 5)
Set 1
Associated arrays
exercise fight gamble future
healthy punch casino past
fitness violent poker always
jogging struggle chance present
Non-associated arrays
gamble casino poker chance
future past always present
exercise healthy fitness jogging
punch fight violent struggle
Set 2
Associated arrays
witch gallery comedy journey
spell sculpture joke travel
magic artist laugh abroad
curse exhibition funny holiday
Non-associated arrays
witch gallery joke holiday
exhibition magic sculpture artist
funny abroad curse comedy
travel laugh journey spell
Appendix 7 Stimulus arrays probed in associated and non-associated conditions of the two word sets in the concrete associates task (Experiment 5)
Set 1
Associated arrays
farm oven sailor shelf
cow apron boat wall
tractor kitchen anchor book
barn soup sea wood
Non-associated arrays
sailor boat anchor sea
shelf wall book wood
farm cow tractor barn
oven apron kitchen soup
Set 2
Associated arrays
monkey airport picture tent
cage plane brush camp
banana suitcase paint woods
zoo ticket canvas fire
Non-associated arrays
tent zoo paint picture
monkey plane airport camp
suitcase canvas woods ticket
brush fire banana cage
Acknowledgements
We wish to thank Dr D Cohen for allowing us to study a patient under his care. We are also indebted to Katie Masters for her kind assistance and for providing us with language assessment information.
References
Allport DA. Distributed memory, modular systems and dysphasia. In: Newman SK, Epstein R, editors. Current perspectives on dysphasia. Edinburgh: Churchill Livingstone; 1985.
Anderson RC, Nagy WE. Word meanings. In: Pearson PD, editor. Comprehension and teaching: research reviews. New York: Longman; 1991.
Breedin SD, Saffran EM, Coslett HB. Reversal of the concreteness effect in a patient with semantic dementia. Cognit Neuropsychol 1994; 11: 617–60.
Brown RW. Linguistic determinism and the part of speech. J Abnorm Soc Psychol 1957; 55: 1–5.
Capitani E, Laiacona M, Mahon B, Caramazza A. What are the facts of semantic category-specific deficits A critical review of the clinical evidence. Cognit Neuropsychol 2003; 20: 213–61.
Coltheart M. The MRC Psycholinguistic Database. Q J Exp Psychol 1981; 33A: 497–505.
Crutch SJ, Warrington EK. Refractory dyslexia: evidence of multiple task-specific phonological output stores. Brain 2001; 124: 1533–43.
Crutch SJ, Warrington EK. Spatial coding of semantic information: knowledge of country and city names depends upon their geographical proximity. Brain 2003a; 126: 1821–9.
Crutch SJ, Warrington EK. Preservation of propositional speech in a pure anomic: the importance of an abstract vocabulary. Neurocase 2003b; 9: 465–81.
Crutch SJ, Warrington EK. The semantic organization of proper nouns: the case of people and brand names. Neuropsychologia 2004a; 42: 584–96.
Crutch SJ, Warrington EK. Gradients of semantic relatedness and their contrasting explanations in refractory access and storage semantic impairments. Cognit Neuropsych. In press, 2005.
Dunn LM, Dunn LM, Whetton C, Pintilie D. The British Picture Vocabulary Scale. Windsor (UK): NFER-Nelson; 1982.
Farah MJ, McClelland JL, A Computational model of Semantic memory impairment: Modality specificity and emergent category specificity, Exp. Psychol Gen. 1991; 120: 339–57.
Forde EME, Humphreys GW. Refractory semantics in global aphasia: on semantic organization and the access-storage distinction in neuropsychology. Memory 1995; 3: 265–307.
Franklin S, Howard D, Patterson K. Abstract word anomia. Cognit Neuropsychol 1995; 12: 549–66.
Gentner D. Why nouns are learned before verbs: linguistic relativity versus natural partitioning. Center for the Study of Reading: Technical Report No. 257. Urbana (IL): University of Illinois at Urbana-Champaign; 1982.
Gotts SJ, Plaut DC. The impact of synaptic depression following brain damage: a connectionist account of ‘access/refractory’ and ‘degraded store’ semantic impairments. Cognit Affect Behav Neurosci 2002; 2: 187–213.
Hillis AE, Rapp BC, Romani C, Caramazza A. Selective impairment of semantics in lexical processing. Cognit Neuropsychol 1990; 7: 191–243.
Hodges JR, Patterson K, Tyler LK. Loss of semantic memory: implications for the modularity of mind. Cognit Neuropsychol 1994; 11: 505–42.
Hodges JR, Graham N, Patterson K. Charting the progression in semantic dementia: implications for the organization of semantic memory. Memory 1995; 3: 463–95.
Howard D, Patterson K. The Pyramids and Palm Trees Test. Bury St Edmunds: Thames Valley Test Company; 1992.
James CT. The role of semantic information in lexical decisions. J Exp Psychol: Hum Percept Perform 1975; 104: 130–6.
Kaplan E, Goodglass H, Weineraab S. Boston naming Test. Media, PA: Williams and Wilkins; 1983.
Landau B, Jones S, Smith L. Perception, ontology and naming in young children. Commentary on Soja, Carey and Spelte. Cognition 1992; 43: 85–91.
Landau B, Smith L, Jones S. Object perception and object naming in early development. Trends in Cognative Sciences 1998; 2: 19–24
Marshall JC, Newcombe F. Patterns of paralexia: a psycholinguistic approach. J Psycholinguist Res 1973; 2: 175–99.
Marshall J, Pring T, Chiat S, Robson J. When ottoman is easier than chair: an inverse frequency effect in jorgan aphasia. Cortex 2001; 37: 33–53.
McCarthy RA, Kartsounis LD. Wobbly words: refractory anomia with preserved semantics. Neurocase 2000; 6: 487–97.
McClelland JL, Rogers TT. The parallel distributed processing approach to semantic cognitions. Nat Rev Neurosci. 2003; 4: 310–22.
McKenna P, Warrington EK. Category-specific naming preservation: a single case study. J Neurol Neurosurg Psychiatry 1978; 41: 571–4.
Nickels L, Howard D. Phonological errors in aphasic naming: comprehension, monitoring and lexicality. Cortex 1995; 31: 209–37.
Paivio A, Yuille C, Madigan SA. Concreteness, imagery and meaningfulness values for 925 nouns. J Exp Psychol 1968; 76 (Monograph Suppl): 1–25.
Paivio A. Imagery and verbal processes. New York: Holt, Rinehart & Winston; 1971.
Paivio A. Mental representations: a dual coding approach. Oxford: Oxford University Press; 1986.
Plaut DC, Shallice T. Effects of word abstractness in a connectionist model of deep dyslexia. In: Proceedings of the 13th Annual Meeting of the Cognitive Science Society. Chicago: Cognitive Science Society; 1991. p. 73–8.
Plaut DC, Shallice T. Deep dyslexia: a case study of connectionist neuropsychology. Cognit Neuropsychol 1993; 10: 377–500.
Rodd J, Gaskell G, Marslen-Wilson W. Making sense of semantic ambiguity: semantic competition in lexical access. J Mem Lang 2002; 46: 245–66.
Rodd J, Gaskell G, Marslen-Wilson W. Modelling the effects of semantic ambiguity in word recognition. Cognit Sci 2004; 28: 89–104.
Schwanenflugel PJ, Shoben EJ. Differential context effects in the comprehension of abstract and concrete verbal materials. J Exp Psychol Learn Mem Cogn 1983; 9: 82–102.
Schwanenflugel PJ, Stowe RW. Context availability and the processing of abstract and concrete words. Read Res Q 1989; 24: 114–26.
Schwanenflugel PJ, Harnishfeger KK, Stowe RW. Context availability and lexical decisions for abstract and concrete words. J Mem Lang 1988; 27: 499–520.
Sirigu A, Duhamel JR, Poncet M. The role of sensorimotor experience in object recognition. A case of multimodal agnosia. Brain 1991; 114: 2555–73.
Warrington EK. The selective impairment of semantic memory. Q J Exp Psychol 1975; 27: 635–57.
Warrington EK. Neuropsychological studies of verbal semantic systems. Philos Trans R Soc Lond B Biol Sci 1981; 295: 411–23.
Warrington EK, McCarthy RA. Category specific access dysphasia. Brain 1983; 106: 859–78.
Warrington EK, Cipolotti L. Word comprehension. The distinction between refractory and storage impairments. Brain 1996; 119: 611–25.
Warrington EK, Crutch SJ. A circumscribed refractory access disorder: a verbal semantic impairment sparing visual semantics. Cognit Neuropsychol 2004; 21: 299–315.
Warrington EK, Shallice T. Category specific semantic impairments. Brain 1984; 107: 829–54.(Sebastian J. Crutch, and Elizabeth K. Wa)
Summary
The architecture supporting our conceptual knowledge of abstract words has remained almost entirely unexplored. By contrast, a vast neuropsychological, neurolinguistic and neuroimaging literature has addressed questions relating to the structure of the semantic system underpinning our knowledge of concrete items (e.g. artefacts and animals). In the context of semantic refractory access dysphasia, a series of experiments exploring and comparing abstract and concrete word comprehesion are described. We demonstrate that semantically associated abstract words reliably interfere with one another significantly more than semantically synonymous abstract words, while concrete words show the reverse pattern. We report the first evidence that abstract and concrete word meanings are based in representational systems that have qualitatively different properties. More specifically, we show that abstract concepts, but not concrete concepts, are represented in an associative neural network. Furthermore, our patient was found to have significantly greater difficulty in identifying high frequency than low frequency abstract words. This observation constitutes the first evidence of an inverse word frequency effect. Our results challenge the generality of many existing models of human conceptual knowledge, which derive their structure from experimental findings in the concrete domain alone.
Key Words: semantic refractory access dysphasia; conceptual knowledge; abstract concepts; concrete concepts
Abbreviations: RSI = response-stimulus interval
Introduction
In almost 30 years since the first publication detailing a selective impairment of semantic memory, research has revealed an ever more fine-grain categorical organization of human knowledge for concrete concepts (for a review, see Capitani et al., 2003). Over the same period, there has been a great paucity of comparable investigations of abstract conceptual knowledge. This is despite the fact that, in the original report of semantic memory deficits, the performance of two patients tested using the same stimuli revealed a double dissociation between concrete and abstract word comprehension (Warrington, 1975). The more unusual pattern of a selective impairment of concrete word knowledge with relatively preserved abstract word knowledge was subsequently replicated (e.g. Warrington and Shallice, 1984; Sirigu et al., 1991; Breedin et al., 1994). The paucity of experimental work on the abstract portion of our vocabulary has not gone unnoticed though, with Breedin et al. (1994) noting that:
‘the present results also point to a gap in the psychological literature on semantic representation, which has little to say about the manner in which abstract concepts are represented’ (p. 649).
The importance of the distinction between concrete and abstract concepts has long been recognized in a number of cognitive domains. For example, the role of item concreteness has been stressed in the context of normal adults' recall, recognition and lexicial decision abilities (Paivio, 1971; James, 1975), neurologically damaged patients' reading skills (e.g. deep dyslexics; Marshall and Newcombe, 1973), and children's vocabulary development (Brown, 1957; Gentner, 1982). Typically, subjects show a processing advantage for concrete words (e.g. leaf, house, sword) over abstract words (e.g. pity, victory, deceit). This superiority has been attributed variously to: (i) abstract words lacking the direct sensory referents of concrete words (Paivio, 1986); (ii) greater availability of contextual information in the knowledge base for concrete words (Schwanenflugel and Shoben, 1983); and (iii) concrete words being supported by more semantic features than abstract words (Plaut and Shallice, 1991, 1993). One common theme in these theories is that a quantitative distinction is drawn between concrete and abstract concepts, with concrete items more strongly represented. By contrast, we argue here that the fundamental distinction is rooted in qualitatively different principles of organization.
Theories of the representational framework underpinning conceptual knowledge have largely been restricted to the domain of concrete items, frequently being stimulated by examinations of the phenomenon of category-specific semantic deficits. Indeed, it was the observation of the selective inability of some patients to retrieve semantic information concerning certain classes of items which led Warrington (1981) to posit that semantic space had a fine-grain categorical organization. It should be noted, however, that in advocating a role for categorical information, it was specifically predicted that other types of information such as contextual association would be shown not to be powerful organizational principles:
‘I have argued that the verbal semantic systems are categorical in organization ... In putting such an argument are we rejecting any alternative model of semantic organization Yes, I think so. Such a formulation is almost certainly incompatible with notions that consider verbal semantic organization necessarily to be in the form of a network of words related by associations. To be more concrete: lion/danger or horseshoe/luck would be close in an associative network organization but distant in a categorical organization’ (Warrington, 1981; p. 422).
Nonetheless, other researchers have proposed that concepts are underpinned by an associative network in the sense that category-specific effects emerge after regional damage to an associated set of semantic features (e.g. Allport, 1985; Farah and McClelland, 1991; Hodges et al., 1994, 1995; McClelland and Rogers, 2003). Although we address the debate surrounding concrete concept organization by explicitly testing Warrington's (1981) prediction, the focus of the current experimental investigations is upon the less frequently studied knowledge of abstract concepts. In particular, we compare whether abstract and concrete word meanings are based in representational systems that have the same or different properties.
One approach to studying the properties of the semantic system underlying concrete concepts has been to examine the comprehension performance of patients with semantic refractory access dysphasia. Refractoriness is an observable, behavioural phenomenon, which can be defined by the influence of temporal factors upon performance of a cognitive task. More specifically, the performance of individuals with semantic refractory access disorders is facilitated if the delay between each response and the presentation of the subsequent stimulus is increased (response-stimulus interval; RSI).
The problem underlying refractoriness has been described as:
the reduction in the ability to utilize the system for a period of time following activationWarrington and McCarthy, 1983; p. 874.
Refractoriness has a number of behavioural consequences. First, refractoriness causes response inconsistency; with repetitive testing of a set of stimuli, identification of items may be achieved once but become unreliable thereafter. Secondly, in patients with refractory access disorders, the strong and reliable effects of word frequency are often reduced or absent. Finally, refractoriness leads to a sensitivity to the semantic relatedness of test stimuli. Refractoriness can be observed to affect not only individual concepts, but also previously unprobed concepts which are represented close in neural space. It is the last of these three consequences that has been exploited to reveal a number of very fine-grain organizational principles in the domain of concrete vocabulary (e.g. types of man-made artefacts: Warrington and Crutch, 2004) and in the domain of proper names (e.g. cities: Crutch and Warrington, 2003a).
At the physiological level, the refractory access phenomena may be caused by damage to the neuromodulatory systems which mediate neural activation levels compared with a relative sparing of the neurons which directly encode semantic information (Gotts and Plaut, 2002). This neuromodulatory damage results in excessive neuronal depression and a refractory period during which subsequent neural firing is blocked or reduced.
Refractory access disorders have also been observed to affect non-semantic components of the language system (e.g. refractory anomia: McCarthy and Kartsounis, 2000; refractory dyslexia: Crutch and Warrington, 2001). Consequently, in studying comprehension abilities in patients with refractory access conditions it is essential to verify that the refractory processes held to affect behaviour are operating at a semantic and not pre- or post-semantic level. In the case of the patient reported in this paper, A.Z., this was achieved by examining her ability to match spoken words to pictures grouped in arrays by semantic relationship (e.g. goat, deer, bear) compared with phonological relationship (i.e. rhymes such as goat, boat, coat). A.Z. made significantly more errors in the semantic than phonological condition, indicating that at least a major component of her comprehension deficit results from an impairment at the level of semantic processing (Warrington and Crutch, 2004). A semantic locus for A.Z.'s refractory access deficit is further corroborated by the presence of semantic distance effects on word–picture matching tasks when the identity of a set of items is probed under both semantically close (i.e. all potential responses drawn from the same category) and semantically distant array conditions (i.e. all potential responses drawn from different categories). Most importantly in relation to the experiments reported in this paper, equivalent semantic distance effects are also observed on tasks conducted entirely within the verbal domain, namely spoken word–written word matching. These effects provide strong evidence that the matching procedure is mediated at least in part at a semantic level (e.g. Crutch and Warrington, 2003a, 2004a).
It is this demonstration of semantic mediation on verbal matching tasks that opens the way for exploration of the domain of abstract conceptual knowledge. Indeed one reason for the dearth of abstract word experimentation is that, by their nature, low concreteness concepts have low imageability and hence cannot easily be submitted to investigation using traditional neuropsychological techniques such as picture naming or picture identification. Furthermore, the use of spoken word to written word matching tasks is unsuitable for the majority of patients with a central semantic deficit (e.g. patients with semantic dementia) because the task might still be performed without recourse to semantic mediation via a process of print–sound correspondence. As noted though, these techniques can and have been used to explore fine-grain categories within a wide-ranging vocabulary in patients with a refractory access disorder of semantic processing. Here the technique is applied for the first time to abstract concepts.
a more practical perspective, semantic distance effects have typically been measured in the concrete domain by comparing patients' abilities to identify target items from among distractor items drawn from the same semantic category (e.g. celery, radish, leek, onion) as opposed to distractors drawn from different semantic categories (e.g. celery, rhino, flask, helicopter). In the case of patients with a semantic refractory access deficit, refractoriness is observed to build up much more quickly in the former than in the latter condition (e.g. Forde and Humphreys, 1995; Warrington and Cipolotti, 1996; Warrington and Crutch, 2004). However, most abstract words are not as easily classified as concrete words. For example, the word ‘dog’ can be classed as an animal (superordinate) and has various exemplars such as ‘labrador’ and ‘alsatian’ (subordinates), but the word ‘luck’ cannot easily be categorized in the same manner. Many abstract words do, however, have synonyms, i.e. other abstract words which have essentially the same meaning (e.g. reluctant, unwilling, averse). Indeed, in the studies reported here we use synonymy as a marker of semantic similarity among abstract words.
In the context of a semantic refractory access disorder, this series of experiments attempts to measure the relative decrement in comprehension accuracy as influenced by similarity and association of meaning. Thereby, we attempt to compare directly semantic similarity and semantic association as competing principles of organization of both abstract and concrete word semantics. In doing so, we document the effects of abstractness and frequency upon the comprehension abilities of our patient.
Case report
A.Z. is a 73-year-old right handed lady who became globally aphasic following a left middle cerebral artery haemorrhage in January 2002. A CT scan demonstrated extensive damage to the left middle cerebral territory with left parietal, temporal and posterior frontal involvement. A.Z.'s spontaneous speech was fluent yet jargonistic with frequent semantic and phonemic errors. She also demonstrated a gross and pervasive impairment of her reading, writing and naming skills. In particular, she was unable to read high frequency monosyllabic words, could not write her own name and named only two items from the Boston Naming Test.
With regard to her comprehension skills, A.Z. exhibited a mild static semantic impairment. On the short form of the British Picture Vocabulary Scale (Dunn et al., 1982), she scored 22 out of 32 on oral presentation and 23 out of 32 with written stimuli. Furthermore, on the Pyramid and Palm Trees Test (Howard and Patterson, 1992), A.Z. gained the weak score of 38 out of 52 on both the word–word matching and picture–picture matching versions of the task. Comprehension performance on a high frequency word-picture matching task revealed no evidence of a category-specific semantic deficit (McKenna and Warrington, 1978).
Therefore, despite having no viable expressive language skills, systematic investigation of A.Z.'s comprehension abilities remained possible with the use of simple written word or picture pointing tasks. These techniques have previously been applied to three aspects of her conceptual knowledge, namely place names (Crutch and Warrington, 2003a), person names (Crutch and Warrington, 2004) and modality-specificity (Warrington and Crutch, 2004). In this paper, however, we tackle a much larger and arguably more important component of conceptual knowledge, namely the realm of abstract words. The investigations reported here were conducted between June 2003 and January 2004, and were designed to document and compare A.Z.'s comprehension of abstract and concrete vocabularies.
Experiment 1: Frequency and abstractness effects in abstract word identification
Words are often classified as ‘abstract’ or ‘concrete’, particularly in the context of neuropsychological experimentation. However, abstractness/concreteness may in fact constitute a continuum ranging between highly abstract concepts (e.g. tendency) and highly concrete concepts (e.g. spoon). If this were the case, then a number of reports of concreteness effects in patients may reflect a greater expression or an exacerbation of a pre-existing abstractness gradient. In this experiment, we explored the influence of the degree of abstractness and word frequency upon A.Z.'s word identification skills. More specifically, abstractness was manipulated without resorting to stimuli drawn from the concrete extreme of any potential continuum; only words traditionally classified as abstract were included in the experiment.
Methods
The stimuli consisted of two sets of 144 abstract written words with frequency ratings and abstractness ratings drawn from the MRC Psycholinguistic Database (Coltheart, 1981; see Appendix 1). All words were between five and eight letters in length. The abstract words in both sets were divided equally into a matrix of three levels of frequency (low, middle and high frequency) x three levels of abstractness (highly abstract, moderately abstract and mildly abstract). For the frequency ratings, there was no overlap between the matrix cells, with each frequency band defined by a strict range of Thorndike–Lorge ratings (low: 1–15; middle: 100–200; high: >300). Similarly for the abstractness ratings, there was no overlap between the ratings that defined each band (very low: 200–300; low: 300–400; mild: 400–500). The 16 items in each matrix cell of each set were arranged into four written word arrays of four items. The experiment was conducted as a spoken word–written word matching task, so that for each group of four arrays, one item was probed from each array in sequence until every item had been probed once. A.Z. was required to point to the target written word following each spoken word stimulus. The experiment was conducted at a natural presentation rate with a stimulus-response interval of 1 s. For the first set of words, low, middle and high frequency arrays were tested in order for the very low abstractness words, and latterly for the low and mildly low words. For the second set of words, the identical procedure was used except that arrays corresponding to each cell of the matrix were presented in the reverse order. More specifically, the high frequency/mild abstractness cell was tested first and the low frequency/very low abstractness cell was tested last.
Results
The percentage correct responses for each cell of the word matrix (both stimulus sets combined) together with the percentage total correct responses for each frequency and abstractness band are shown in Table 1. A 2 test for trend of odds revealed that A.Z. showed significant trends to better performance not only with less abstract words [2(1) = 8.95, P < 0.01] but also with lower frequency words [2(1) = 4.56, P < 0.05].
Comment
These results demonstrate that A.Z. shows an increase in impairment in written word identification the greater the abstractness of the target word. However, the pattern of performance is atypical in the sense that the results reported here document for the first time the reversal of the typical word frequency effect. Patients with semantic refractory access disorders have previously been shown to be relatively insensitive to the effects of word frequency (i.e. the absence of a frequency effect), but have never before exhibited an actual reversal of the tendency. The superiority of low frequency word performance is particularly striking, as an advantage for high frequency words is almost ubiquitous in clinical and experimental studies of all components of the language processing system.
Experiment 2: The effect of temporal factors upon abstract word comprehension
In a number of previous word comprehension studies, we have demonstrated that A.Z. exhibits the core features of a refractory access deficit. In particular, A.Z. shows a sensitivity to temporal factors, with performance being facilitated by increasing the length of the interval between making each response and the presentation of the subsequent stimulus (RSI). However, these demonstrations have been conducted in the domains of concrete concepts and proper nouns. Therefore, in order to study semantic distance effects in A.Z.'s comprehension of abstract words, it was considered necessary to establish whether the influence of temporal factors was also apparent in this domain of conceptual knowledge.
Methods
The stimuli consisted of 20 written abstract words, arranged into five arrays each containing four items (e.g. wait, task, ache, stiff; see Appendix 2). For each array, the name of every written word was spoken by the examiner four times in a pseudo-random order. Following the presentation of each spoken word target, A.Z. was requested to point to the corresponding written word in the array. This procedure was conducted twice for each array, once under a fast presentation rate (1 s RSI) and once under a slow presentation rate (10 s RSI). The two temporal conditions were alternated in an ABABA BABAB design.
Results
The results for each condition are shown in Fig. 1. A Wilcoxon matched-pairs signed-ranks test showed that A.Z. made significantly more errors in the fast condition than in the slow condition (z = 2.50, P < 0.02; Fig. 1A). Having probed each item four times, an item-specific analysis was conducted to examine on which probe each error occurred. Serial position curves showing the percentage error rate on each probe are given in Fig. 1B. In the fast condition, a Cochran Q test revealed a significant difference in the number of errors made on each probe [Q(3) = 16.80, P < 0.001]. This difference reflects a gradual increase in error rate with successive probes following initially accurate performance. In the slow condition however, A.Z.'s performance was almost at ceiling and there was no significant difference in the number of errors on each probe [Q(3) = 3.62, P > 0.3]. Thus, A.Z.'s ability to comprehend the meaning of abstract words is characterized by at least two of the core features of a refractory access disorder, namely sensitivity to temporal factors and inconsistency.
Experiment 3: The influence of semantic similarity upon abstract adjective and verb comprehension
The exploration of semantic organization in the context of refractory access dysphasia has typically depended upon the manipulation of semantic distance effects. In particular, it has been shown that, despite accurate performance initially, refractoriness builds up more quickly between semantically related concepts than semantically unrelated concepts leading to a higher overall error rate. Semantic similarity has been demonstrated to affect concrete word comprehension performance not only in patient A.Z., but also in other semantic refractory access patients (e.g. Forde and Humphreys, 1995; Warrington and Cipolotti, 1996). However, equivalent effects of semantic similarity in the abstract word domain have not previously been investigated. In this experiment, comprehension of two classes of abstract words was investigated, namely adjectives and verbs. Unlike in the concrete domain in which ‘semantic relationship’ is most frequently defined by membership of a common semantic category, few abstract concepts can be classified in the same manner. Consequently, we examined semantic distance among adjectives using synonyms derived from a thesaurus to form semantically similar and semantically dissimilar groups. By contrast, the verbs selected in this task were not synonymous but were chosen because they formed coherent, semantically related clusters of concepts.
Methods
The stimuli consisted of two sets of 16 adjectives and one set of 16 verbs (see Appendix 3). The items within each adjective set were arranged into four semantically related written word arrays (e.g. divine, sacred, religious, holy). The same words were also arranged into four semantically unrelated arrays, each containing one word from each the semantically related arrays (e.g. divine, sparse, furious, sudden). In an identical manner, the items within the verb set were also arranged into both semantically related (e.g. kneel, sit, stand, lie) and semantically unrelated written word arrays (e.g. kneel, sing, bake, drive). The tasks were conducted as spoken word–written word matching tests. Each item in an array was probed four times in a pseudo-random order and a 1 s RSI was adopted. Comprehension of each set of words was tested separately and, within each set, the arrays were presented in an ABBA order.
Results
The percentage correct responses for adjectives (both word sets combined) and verbs under the conditions of semantic relatedness and unrelatedness are given in Fig. 2A. Despite the semantic similarity of words in the related arrays, Wilcoxon matched-pairs ranked-sums tests revealed no difference between A.Z.'s ability to identify adjectives or verbs under the two conditions (adjectives: z = 1.16, P > 0.2; verbs: z = 0.93, P > 0.3). Serial position curves demonstrating the percentage error rate for each probe of adjective and verb items are shown in Fig. 2B. Cochran Q tests revealed that A.Z. showed a significant difference in the number of errors made on successive probes with both the adjectives [related: Q(3) = 29.40, P < 0.001; unrelated: Q(3) = 17.89, P < 0.001] and the verbs [related: Q(3) = 8.48, P < 0.05; unrelated: Q(3) = 10.00, P < 0.02]. Examination of Fig. 2B indicates that these differences reflect increasing error rates with repeated probes following an almost ceiling level of performance on probe 1.
Comment
Overall, the data reported in this experiment suggest that refractoriness does not build up more quickly among semantically related abstract words more than semantically unrelated words. The equivalence of performance on related and unrelated arrays of both adjectives and verbs was unexpected, and very different to the pattern of performance predicted for concrete nouns.
Experiment 4: The influence of semantic similarity upon abstract and concrete word comprehension
In the context of a semantic refractory access dysphasia, semantic distance effects are attributed to the abnormal, deleterious effects which activating a concept has upon other concepts that partly share neural space. Thus, the failure to observe semantic distance effects among abstract words in Experiment 3 was highly surprising, not only because such effects are almost ubiquitous within the concrete domain, but because the observation implies that abstract words with similar meanings do not necessarily share neural space. In this experiment, the influence of semantic similarity upon A.Z.'s comprehension abilities was further investigated by comparing and contrasting abstract and concrete words. Given that there are very few adjectives with high concreteness values, the abstract words chosen for this experiment were mainly nouns and verbs.
Methods
The stimuli consisted of two sets of abstract words (see Appendix 4) and two sets of concrete words (see Appendix 5). Within each set of abstract words, the items were arranged into semantically synonymous written word arrays each containing four abstract synonyms (e.g. deceit, trick, steal, cheat). The same words were also arranged in semantically non-synonymous arrays consisting of one word from each of four synonymous arrays (e.g. deceit, strike, mush, screen). In an identical fashion, the items within each set of concrete words were arranged into four written word arrays in which each item was drawn from the same semantic category (e.g. goose, pigeon, crow, sparrow). The same concrete words were also arranged into four different semantic category arrays containing one item from each of the same category arrays (e.g. goose, melon, pullover, biscuit). All other aspects of the experimental design were identical to the procedure employed in Experiment 3.
Results
The combined percentage correct responses for the synonymous and non-synonymous abstract word conditions are shown in Fig. 3A. A Wilcoxon matched-pairs signed-ranks test revealed no difference in response accuracy between the synonymous and non-synonymous array conditions (z = 0.22, P > 0.8). Serial position curves demonstrating the percentage error rate for each probe of the abstract words are also shown in Fig. 3A. Cochran Q tests revealed that A.Z. showed a significant difference in the number of errors made on successive probes [synonyms: Q(3) = 17.48, P < 0.001; non-synonyms: Q(3) = 12.00, P < 0.01]. Inspection of the serial position curves suggests that these differences reflect an increase in the percentage error rate across consecutive probes of the test stimuli for both synonymous and non-synonymous conditions alike.
The percentage correct responses on the concrete word stimulus sets are shown in Fig. 3B. The results for both word sets were considered together. A Wilcoxon matched-pairs signed-ranks test showed that A.Z. was significantly less accurate at pointing to the target word in same semantic category arrays than different semantic category arrays (z = 3.62, P < 0.001). Serial position curves showing the percentage error rate for successive probes of the test items under both same category and different category conditions are also given in Fig. 3B. Cochran Q tests again demonstrated a significant difference in response accuracy between probes under the same category condition [Q(3) = 14.76, P < 0.01] and a trend towards a similar pattern in the different category condition [Q(3) = 7.43, P < 0.1]. Both these results describe a gradual increase in the error rate across successive probes.
Comment
The evidence presented in this experiment strongly suggests that refractoriness does not build up more quickly among abstract words which have similar meanings than among those which do not. By contrast, significant effects of semantic relatedness were observed in the comprehension of more concrete words. The surprising absence of the semantic relatedness effect among abstract words cannot be attributed to the grammatical class of the abstract word stimuli in question as equivalent null results were obtained using adjectival stimuli as when using noun and verb stimuli.
Experiment 5: The influence of semantic association upon abstract and concrete word comprehension
Having established that abstract word representations are unaffected by semantic similarity (i.e. category) relative to concrete words, we evaluated an alternative potential principle of organization: semantic association. Associated words are those whose meanings are not synonymous but which are often bound together in real world or sentential contexts (e.g. salute, army, general, respect). The significance of associative links between concepts was examined using the same spoken word to written word matching technique described in Experiments 3 and 4.
Methods
The stimuli consisted of two sets of 16 abstract words (see Appendix 6) and two sets of 16 concrete words (see Appendix 7). For all stimulus sets, the items were arranged into arrays each containing four associated but non-synonymous words (e.g. abstract: exercise, healthy, fitness, jogging; concrete: farm, cow, tractor, barn). The same words were also arranged into four non-associated arrays with one word drawn from each associated array (e.g. abstract: exercise, gamble, punch, future; concrete: farm, sailor, shelf, oven). As before, the tasks were conducted as spoken word–written word matching tests. Each item in an array was probed four times in a randomized sequence, using a 1 s RSI. Comprehension of each set of words was tested separately and, within each set, the arrays were administered in an ABBA design.
Results
The percentage correct responses (sets 1 and 2 combined) for the abstract words under both associated and non-associated array conditions are shown in Fig. 3C. A Wilcoxon matched-pairs signed-ranks test showed that A.Z. showed a difference between conditions with significantly worse performance on the associated word arrays than non-associated word arrays (z = 4.23, P < 0.001). Serial position curves showing the percentage error rate across the four probes of each test item are also given in Fig. 3C. Cochran Q tests demonstrated a strong trend towards different accuracy levels between probes in the associated array condition [Q(3) = 6.48, P < 0.1] and a significant difference between probes in the non-associated condition [Q(3) = 15.67, P < 0.01]. Inspection of the serial position curves suggests that both these effects correspond to an increase in the percentage error rate across consecutive probes of the test stimuli.
The percentage correct responses on the concrete word stimulus sets are shown in Fig. 3D. The results for both word sets were considered together. A Wilcoxon matched-pairs signed-ranks test revealed no significant difference between the associated and non-associated conditions (z = 0.23, P > 0.8). As with the abstract word stimuli, serial position curves showing the percentage error rate for successive probes of the test items are also given in Fig. 3D. Cochran Q tests also demonstrated significant differences in response accuracy between probes under both the associated array condition [Q(3) = 15.42, P < 0.01] and the non-associated array condition [Q(3) = 22.83, P < 0.001]. Both these results describe a gradual increase in the error rate across successive probes.
Comment
The significant effect of semantic association upon the comprehension of abstract words makes a striking contrast to the absence of semantic similarity effects in the abstract synonyms tasks. It appears that refractoriness builds up much more easily among associated than synonymous abstract words, implying that the representations underpinning associated abstract words are much closer neural neighbours than the representations which support words which have very similar meanings. In the concrete domain, evidence of the reverse pattern has been documented. The absence of an effect of semantic association upon the comprehension of concrete words stands in contrast to the strong effect of semantic similarity. Thus these results give further strength to the claim that, unlike abstract concepts, concrete concept representations are organized by semantic category and not by semantic association.
Discussion
The observation and manipulation of semantic relatedness effects in patients with semantic refractory access disorders enable the investigation of principles governing the organization of semantic memory. It is held that refractoriness may be observed between related concepts because the activation of one representation leads to the partial activation of other representations that share neural space. In previous reports of our patient A.Z., we have described a number of such effects in the conceptual domains of concrete items and proper nouns (Crutch and Warrington, 2003a, 2004, 2005; Warrington and Crutch, 2004). In this paper, we have extended this approach to the rarely studied domain of abstract words. Sets of synonymous abstract words were selected to be analogous to the semantically related categorical groupings of concrete items employed in earlier studies. However, to our great surprise, we were entirely unable to detect semantic relatedness effects among synonymous abstract word stimuli. This unexpected absence of a relatedness effect was documented for several sets of stimuli. It was demonstrated that adjectives (Experiment 3), verbs (Experiment 3) and also mixed noun and verb parts of speech arrays (Experiment 4) were all equally immune to the effect. By contrast, using an identical spoken word–written word matching procedure, strong semantic distance effects continued to be observed among concrete words (Experiment 4). The unpredicted null results led us to consider alternative forms of relationship between abstract words and, in particular, contextual associative relationships. On a task comparing A.Z.'s ability to identify associated and non-associated abstract words, A.Z. showed significantly worse performance in the associated condition (Experiment 5). However, remarkably no effect of semantic association was detected on an equivalent task involving concrete words (Experiment 5). Thus, we have documented evidence of a double dissociation between abstract words which show an effect of semantic contextual association but not semantic similarity, and concrete words which show an effect of semantic similarity but not semantic association. On the basis of these contrasting patterns of results, we would claim that abstract concepts are represented in an associative neural network whereas concrete concepts have a categorical organization. Indeed, while some researchers have claimed that concrete concepts are imbedded in a hierarchy of superordinate and subordinate representations, this principled organization appears not to apply to a more abstract vocabulary. We would predict that these contrasting organizations must be reflected in the neural substrate of conceptual knowledge.
In this discussion, we first consider existing views in the cognitive neuroscience literature upon the difference between abstract and concrete concepts before considering the implications of our own results. In both of these considerations, we make a clear distinction between theories of how the abstract–concrete difference is acquired and theories of how that difference is represented.
It has been claimed that the abstract–concrete difference reflects the manner in which these concepts are acquired and the roles they play in language (Breedin et al., 1994). Experience of the perceptual features of objects via our five sensory channels appears to play a key role in the acquisition of concrete concepts (e.g. Landau et al., 1992, 1998). Abstract concepts, however, may be acquired in the context of language without any direct perceptual input. As a consequence, Breedin et al. (1994) suggested that: ‘exposure to multiple sentence contexts plays an important role in the development of representations for abstract concepts’ (p. 650).
Consistent with this viewpoint, abstract words have been claimed to have richer associations, and indeed more meanings, than concrete words (Paivio et al., 1968; Schwanenflugel and Shoben, 1983). In particular, the meaning of concrete words (e.g. horse) tends to be constant across most contexts, whereas abstract words (e.g. appeal) often have related but distinct meanings depending on the context in which they are being used. Experimental support from this notion comes from Schwanenflugel and colleagues who in a series of timed comprehension, naming and meaningfulness judgement tasks showed that the provision of sentence or paragraph contexts eliminated the traditional bias toward faster responses with concrete items (Schwanenflugel and Shoben, 1983; Schwanenflugel et al., 1988, Schwanenflugel and Stowe, 1989).
The question of whether abstract words are represented in a qualitatively or quantitatively different way from concrete words has been considered as an explanation of some abstractness effects (e.g. abstract word anomia; Franklin et al., 1995). However, theoretical accounts of conceptual organization have tended to either concentrate upon quantitative differences between the two types of words or to only model concrete concepts to the exclusion of abstract knowledge. For example, Plaut and Shallice (1993) hypothesized that the difference was a reflection of concrete words being supported by a greater number of semantic features. To examine this notion, they developed a connectionist model of oral reading via semantics in which lesions at any locus produced the typical concreteness effect (i.e. more errors for abstract words). However, this result would appear to be unsurprising given that an a priori decision was taken to train the model on a greater number of semantic features for concrete than abstract terms. By contrast, the model of Allport (1985) only considers meaning for imageable words, positing the existence of attribute domains which correspond to different sensory properties. For those concrete concepts for which the model does offer a system of organization, the underlying representational framework is held to be associative in nature, with connections between distributed features according to their co-occurrence.
To our knowledge, only one previous report has claimed, as we do, that the representational difference between abstract and concrete words might be qualitative rather than quantitative (Breedin et al., 1994). However, these researchers characterized the difference in terms of discrepancy in the number of perceptual and functional features required to distinguish abstract and concrete concepts. Our results suggest that irrespective of the type of information which constitutes abstract and concrete concepts, a much more fundamental difference exists in the actual architecture of those representations. The evidence that the spread of refractoriness between concrete concepts is exacerbated by semantic similarity but not semantic association confirms Warrington's prediction that in the domain of concrete concepts, the primary organizational principle is categorical and not associative (Warrington, 1981). By contrast, it appears that the representations of abstract concepts are predominantly linked to other associated rather than semantically similar items. This conclusion is compatible with a wide range of previous observations about the effects of word concreteness, particularly those of Schwanenflugel and colleagues detailed above. We would regard the facilitatory effect of sentential contexts as a further example of the importance of the associative links between abstract concepts. However, we should be careful to note that, given the probable existence of a continuum between abstractness and concreteness, we suggest this associative/categorical dichotomy is relative rather than absolute. More specifically, we predict that middle concreteness items (e.g. nurse, chemistry) have both associative and categorical connections, in more equal proportions than concepts at either end of the spectrum. It is difficult to envisage how existing models of conceptual knowledge, which have mainly been derived from work on category-specific deficits of concrete item knowledge, can be extended to incorporate our findings on abstract words. Essentially, our findings suggest that attempting to model conceptual knowledge within a unitary system based on a single set of network principles is over simplistic.
With regard to the development of different representational frameworks for concrete and abstract concepts, we agree with Breedin et al. (1994) that multiple contexts may play a major role. It is notable that in a dictionary, concrete items typically have only a few fixed meanings (e.g. tomato occurs in the context of gardening and eating). On the other hand, dictionaries tend to describe abstract words as having multiple meanings depending upon context (e.g. the word ‘set’ has >30 transitive and intransitive senses). Furthermore, the psycholinguistic literature makes a clear distinction between two types of ‘multiple meaning’, namely homonyms which have clearly distinct, unrelated meanings (e.g. ‘clog’ = Dutch footwear, to block up, etc.) and polysemous words which have multiple, semantically related senses (e.g. ‘pain’ = punishment, grief, to cause distress, etc.). Recent evidence from lexical decision tasks in normal subjects has demonstrated the importance of this latter distinction by showing a processing disadvantage for words which have multiple unrelated meanings, but a processing advantage for words which have multiple related senses (Rodd et al., 2002, 2004). However, it should be noted that some researchers have suggested that all words are polysemous to some extent, because their precise meanings change in every unique context in which they appear (Anderson and Nagy, 1991). We would argue that an associative network may have a greater capacity to represent the multiple meanings of abstract words than a more formal, rigid categorical network. Such a representational framework also possesses the flexibility to capture the natural variability in the semantics of both words with unrelated meanings and words with related senses.
One further finding which is of particular note is that our patient A.Z. not only exhibited a commonly observed concreteness effect on written word identification tasks, but also showed some evidence of an inverse frequency effect (Experiment 1). An absence of an advantage for high frequency words has been previously reported in several studies of both refractory access patients (e.g. Warrington and Cipolotti, 1996) and patients with static storage impairments of various components of the language system (e.g. Hillis et al., 1990; Nickels and Howard, 1995). Indeed, in another study of our patient A.Z., frequency effects were noted to be absent on a spoken word–picture matching task (Crutch and Warrington, 2004b). However, to the best of our knowledge, A.Z. is the first patient on record to demonstrate significantly better performance with less commonly occurring stimuli. One previous patient exhibiting a pre-disposition for low frequency word production has been reported, but this was in the context of a classical jargon aphasia in which the nature and information processing locus of the impairment was unclear (Marshall et al., 2001). The absence of frequency effects in refractory access patients has been attributed to an impairment of the neuromodulatory system leading to higher than normal levels of synaptic depression following the activation of a representation. Frequency effects are reduced or absent because synaptic depression is greater when activity levels are higher, as in the case of high frequency items (Gotts and Plaut, 2002). We would claim that this explanation can be extended to an actual inversion of the frequency effect; using very low and very high frequency stimuli as in Experiment 1 may generate a sufficiently large activity discrepancy as to generate significantly more synaptic depression following high frequency than low frequency word identification. It should again be stressed that the inversion of the frequency effect in A.Z. is not necessarily reliable and is very much stimulus-dependent. On this account, an absence rather than an inversion of the frequency effect was observed in our earlier study of the same patient (Crutch and Warrington, 2004b) because the (pictureable) stimuli invoked a smaller frequency range.
To summarize, we have reported a series of experiments which indicate that abstract conceptual knowledge relies upon qualitatively different representations than concrete conceptual knowledge. We would suggest that a deeper understanding of abstract word semantics and their relationship to other aspects of the language system is long overdue, particularly bearing in mind the importance of an abstract vocabulary to our ability to communicate and produce fluent propositional speech (Crutch and Warrington, 2003b). Overall, studies of concrete word representations have yielded valuable insights into the structure of human semantic memory, but it is important to bear in mind that they concentrate upon only a fraction of our total vocabulary.
Appendix 1 Stimuli contained in both Set A and Set B of the abstractnessxfrequency matrix task (Experiment 1)
Very low abstractness words (concreteness rating 200–300)
Low frequency Middle frequency High frequency
Set A Set B Set A Set B Set A Set B
avowal dotage concert approval advice respect
excise impunity motive misery emotion delight
truism throe boast honesty blame attitude
feint pacifism tendency anxiety ideal value
savant infinity loyalty devotion pride variety
levity clemency virtue economy quality trust
enigma sagacity distress sorrow instance opinion
alacrity jeopardy wisdom patience memory spite
analogy atrocity mercy extreme purpose success
overtone reprisal magic despair truth spirit
inaction sanctity upset estimate effort manner
heroism protocol talent nonsense guess effect
alias craven dread capacity feeling cause
unbelief betrayal impulse reality chance least
impetus heresy neglect extent while might
bravery aptitude dignity mystery other thought
Low abstractness words (concreteness rating 300–400)
Low frequency Middle frequency High frequency
Set A Set B Set A Set B Set A Set B
emission stipend culture insult profit claim
calculus pommel retreat skill charm distance
prefix pronoun humanity rescue fortune century
tenure perjury welfare threat strain occasion
stoic sequel flutter session fashion protest
regency axiom crisis amateur strength pleasure
sobriety legality victory transfer relief peace
jargon truce portion menace amount health
adultery carnage incline comedy beauty position
bequest rigidity bargain grief result marriage
context velocity proof angel number dream
acidity edict poverty division problem month
interim deity drama research reason sense
foible mirage budget liberty answer trouble
dogma equity incident reaction moment minute
mutiny agility genius reward right business
Mild abstractness words (concreteness rating 400–500)
Low frequency Middle frequency High frequency
Set A Set B Set A Set B Set A Set B
buffoon oxide shriek footstep speech leader
zenith curfew comrade craft patient degree
vitriol molecule fairy ceremony murder object
pestle abbess embrace grate income accident
galaxy mermaid saint rebel shadow youth
abyss fissure stalk deposit breath design
dynasty glutton adult glitter whisper meeting
venom ignition pioneer resort report travel
cuisine psalm scent autumn summer general
mileage donor device disaster notice enemy
alphabet citation signal lecture figure group
segment mosque groan exhaust party fight
shale maize angle damage matter story
quake rogue route bunch laugh country
brawl microbe holiday shrug evening voice
filth sedative interior payment night friend
Appendix 2 Stimulus arrays tested under both fast and slow presentation conditions (Experiment 2)
rest ache tough work great
job tall steep pain task
firm chore wait high hard
sore halt hurt stiff stop
Appendix 3 Adjective and verb stimulus arrays used in Experiment 3
Adjectives (Set 1)
Synonymous arrays
divine scant enraged sudden
sacred bare furious abrupt
religious sparse annoyed rash
holy deficient angry unexpected
Non-synonymous arrays
scant furious sparse angry
sacred rash holy sudden
annoyed deficient enraged bare
unexpected divine abrupt religious
Adjectives (Set 2)
Synonymous arrays
hurried silent stupid glad
fast quiet dense happy
quick peaceful thick cheerful
speedy still foolish merry
Non-synonymous arrays
merry silent dense speedy
foolish stupid quiet thick
quick glad happy cheerful
still hurried fast peaceful
Verbs
Semantically related arrays
sing kneel bake drive
shout sit cook fly
laugh stand boil sail
talk lie fry ride
Semantically unrelated arrays
sing shout stand ride
kneel sit fry talk
bake cook laugh boil
drive fly sail lie
Appendix 4 Stimulus arrays probed in synonymous and non-synonymous conditions of the abstract synonym task (Experiment 4)
Set 1
Synonymous arrays
boil look gale clean eat
heat peek wind neat chew
cook glance storm pure bite
fry see breeze fresh taste
Non-synonymous arrays
boil heat neat cook see
look peek chew storm taste
gale wind fry pure fresh
clean eat glance bite breeze
Set 2
Synonymous arrays
beat mask cheat soft
knock cloak trick mush
thump hide steal pulp
strike screen deceit smooth
Non-synonymous arrays
beat smooth thump screen
cloak mask hide strike
cheat trick steal deceit
soft knock pulp mush
Appendix 5 Stimulus arrays probed in same category and different category conditions of the two word sets in the concrete categories task (Experiment 4)
Set 1
Same category arrays
goose cardigan pizza melon
crow jacket cake grapes
sparrow blouse bread peach
pigeon pullover biscuit banana
Different category arrays
goose banana pigeon cake
melon cardigan bread blouse
pullover pizza jacket crow
biscuit sparrow grapes peach
Set 2
Same category arrays
kettle rabbit leek daffodil
toaster hedgehog carrot tulip
oven squirrel onion rose
fridge badger parsnip pansy
Different category arrays
squirrel badger rabbit tulip
onion rose kettle oven
toaster parsnip daffodil carrot
pansy fridge leek hedgehog
Appendix 6 Stimulus arrays probed in associated and non-associated conditions of the two word sets in the abstract associates task (Experiment 5)
Set 1
Associated arrays
exercise fight gamble future
healthy punch casino past
fitness violent poker always
jogging struggle chance present
Non-associated arrays
gamble casino poker chance
future past always present
exercise healthy fitness jogging
punch fight violent struggle
Set 2
Associated arrays
witch gallery comedy journey
spell sculpture joke travel
magic artist laugh abroad
curse exhibition funny holiday
Non-associated arrays
witch gallery joke holiday
exhibition magic sculpture artist
funny abroad curse comedy
travel laugh journey spell
Appendix 7 Stimulus arrays probed in associated and non-associated conditions of the two word sets in the concrete associates task (Experiment 5)
Set 1
Associated arrays
farm oven sailor shelf
cow apron boat wall
tractor kitchen anchor book
barn soup sea wood
Non-associated arrays
sailor boat anchor sea
shelf wall book wood
farm cow tractor barn
oven apron kitchen soup
Set 2
Associated arrays
monkey airport picture tent
cage plane brush camp
banana suitcase paint woods
zoo ticket canvas fire
Non-associated arrays
tent zoo paint picture
monkey plane airport camp
suitcase canvas woods ticket
brush fire banana cage
Acknowledgements
We wish to thank Dr D Cohen for allowing us to study a patient under his care. We are also indebted to Katie Masters for her kind assistance and for providing us with language assessment information.
References
Allport DA. Distributed memory, modular systems and dysphasia. In: Newman SK, Epstein R, editors. Current perspectives on dysphasia. Edinburgh: Churchill Livingstone; 1985.
Anderson RC, Nagy WE. Word meanings. In: Pearson PD, editor. Comprehension and teaching: research reviews. New York: Longman; 1991.
Breedin SD, Saffran EM, Coslett HB. Reversal of the concreteness effect in a patient with semantic dementia. Cognit Neuropsychol 1994; 11: 617–60.
Brown RW. Linguistic determinism and the part of speech. J Abnorm Soc Psychol 1957; 55: 1–5.
Capitani E, Laiacona M, Mahon B, Caramazza A. What are the facts of semantic category-specific deficits A critical review of the clinical evidence. Cognit Neuropsychol 2003; 20: 213–61.
Coltheart M. The MRC Psycholinguistic Database. Q J Exp Psychol 1981; 33A: 497–505.
Crutch SJ, Warrington EK. Refractory dyslexia: evidence of multiple task-specific phonological output stores. Brain 2001; 124: 1533–43.
Crutch SJ, Warrington EK. Spatial coding of semantic information: knowledge of country and city names depends upon their geographical proximity. Brain 2003a; 126: 1821–9.
Crutch SJ, Warrington EK. Preservation of propositional speech in a pure anomic: the importance of an abstract vocabulary. Neurocase 2003b; 9: 465–81.
Crutch SJ, Warrington EK. The semantic organization of proper nouns: the case of people and brand names. Neuropsychologia 2004a; 42: 584–96.
Crutch SJ, Warrington EK. Gradients of semantic relatedness and their contrasting explanations in refractory access and storage semantic impairments. Cognit Neuropsych. In press, 2005.
Dunn LM, Dunn LM, Whetton C, Pintilie D. The British Picture Vocabulary Scale. Windsor (UK): NFER-Nelson; 1982.
Farah MJ, McClelland JL, A Computational model of Semantic memory impairment: Modality specificity and emergent category specificity, Exp. Psychol Gen. 1991; 120: 339–57.
Forde EME, Humphreys GW. Refractory semantics in global aphasia: on semantic organization and the access-storage distinction in neuropsychology. Memory 1995; 3: 265–307.
Franklin S, Howard D, Patterson K. Abstract word anomia. Cognit Neuropsychol 1995; 12: 549–66.
Gentner D. Why nouns are learned before verbs: linguistic relativity versus natural partitioning. Center for the Study of Reading: Technical Report No. 257. Urbana (IL): University of Illinois at Urbana-Champaign; 1982.
Gotts SJ, Plaut DC. The impact of synaptic depression following brain damage: a connectionist account of ‘access/refractory’ and ‘degraded store’ semantic impairments. Cognit Affect Behav Neurosci 2002; 2: 187–213.
Hillis AE, Rapp BC, Romani C, Caramazza A. Selective impairment of semantics in lexical processing. Cognit Neuropsychol 1990; 7: 191–243.
Hodges JR, Patterson K, Tyler LK. Loss of semantic memory: implications for the modularity of mind. Cognit Neuropsychol 1994; 11: 505–42.
Hodges JR, Graham N, Patterson K. Charting the progression in semantic dementia: implications for the organization of semantic memory. Memory 1995; 3: 463–95.
Howard D, Patterson K. The Pyramids and Palm Trees Test. Bury St Edmunds: Thames Valley Test Company; 1992.
James CT. The role of semantic information in lexical decisions. J Exp Psychol: Hum Percept Perform 1975; 104: 130–6.
Kaplan E, Goodglass H, Weineraab S. Boston naming Test. Media, PA: Williams and Wilkins; 1983.
Landau B, Jones S, Smith L. Perception, ontology and naming in young children. Commentary on Soja, Carey and Spelte. Cognition 1992; 43: 85–91.
Landau B, Smith L, Jones S. Object perception and object naming in early development. Trends in Cognative Sciences 1998; 2: 19–24
Marshall JC, Newcombe F. Patterns of paralexia: a psycholinguistic approach. J Psycholinguist Res 1973; 2: 175–99.
Marshall J, Pring T, Chiat S, Robson J. When ottoman is easier than chair: an inverse frequency effect in jorgan aphasia. Cortex 2001; 37: 33–53.
McCarthy RA, Kartsounis LD. Wobbly words: refractory anomia with preserved semantics. Neurocase 2000; 6: 487–97.
McClelland JL, Rogers TT. The parallel distributed processing approach to semantic cognitions. Nat Rev Neurosci. 2003; 4: 310–22.
McKenna P, Warrington EK. Category-specific naming preservation: a single case study. J Neurol Neurosurg Psychiatry 1978; 41: 571–4.
Nickels L, Howard D. Phonological errors in aphasic naming: comprehension, monitoring and lexicality. Cortex 1995; 31: 209–37.
Paivio A, Yuille C, Madigan SA. Concreteness, imagery and meaningfulness values for 925 nouns. J Exp Psychol 1968; 76 (Monograph Suppl): 1–25.
Paivio A. Imagery and verbal processes. New York: Holt, Rinehart & Winston; 1971.
Paivio A. Mental representations: a dual coding approach. Oxford: Oxford University Press; 1986.
Plaut DC, Shallice T. Effects of word abstractness in a connectionist model of deep dyslexia. In: Proceedings of the 13th Annual Meeting of the Cognitive Science Society. Chicago: Cognitive Science Society; 1991. p. 73–8.
Plaut DC, Shallice T. Deep dyslexia: a case study of connectionist neuropsychology. Cognit Neuropsychol 1993; 10: 377–500.
Rodd J, Gaskell G, Marslen-Wilson W. Making sense of semantic ambiguity: semantic competition in lexical access. J Mem Lang 2002; 46: 245–66.
Rodd J, Gaskell G, Marslen-Wilson W. Modelling the effects of semantic ambiguity in word recognition. Cognit Sci 2004; 28: 89–104.
Schwanenflugel PJ, Shoben EJ. Differential context effects in the comprehension of abstract and concrete verbal materials. J Exp Psychol Learn Mem Cogn 1983; 9: 82–102.
Schwanenflugel PJ, Stowe RW. Context availability and the processing of abstract and concrete words. Read Res Q 1989; 24: 114–26.
Schwanenflugel PJ, Harnishfeger KK, Stowe RW. Context availability and lexical decisions for abstract and concrete words. J Mem Lang 1988; 27: 499–520.
Sirigu A, Duhamel JR, Poncet M. The role of sensorimotor experience in object recognition. A case of multimodal agnosia. Brain 1991; 114: 2555–73.
Warrington EK. The selective impairment of semantic memory. Q J Exp Psychol 1975; 27: 635–57.
Warrington EK. Neuropsychological studies of verbal semantic systems. Philos Trans R Soc Lond B Biol Sci 1981; 295: 411–23.
Warrington EK, McCarthy RA. Category specific access dysphasia. Brain 1983; 106: 859–78.
Warrington EK, Cipolotti L. Word comprehension. The distinction between refractory and storage impairments. Brain 1996; 119: 611–25.
Warrington EK, Crutch SJ. A circumscribed refractory access disorder: a verbal semantic impairment sparing visual semantics. Cognit Neuropsychol 2004; 21: 299–315.
Warrington EK, Shallice T. Category specific semantic impairments. Brain 1984; 107: 829–54.(Sebastian J. Crutch, and Elizabeth K. Wa)