Neurocase, 10(5): 353–362, 2004 Copyright © Taylor & Francis Inc. 1355-4795/04/1005–353$16.00 Neurocase
Dissociation of Lexical Syntax and Semantics: Evidence from Focal Cortical Degeneration P. Garrard1, E. Carroll1, D. Vinson2 and G. Vigliocco2 1 Institute of Cognitive Neuroscience, London, UK, 2Language Production Laboratory, Dept Psychology, University College London, UK.
Abstract The question of whether information relevant to meaning (semantics) and structure (syntax) relies on a common language processor or on separate subsystems has proved difficult to address definitively because of the confounds involved in comparing the two types of information. At the sentence level syntactic and semantic judgments make different cognitive demands, while at the single word level, the most commonly used syntactic distinction (between nouns and verbs) is confounded with a fundamental semantic difference (between objects and actions). The present study employs a different syntactic contrast (between count nouns and mass nouns), which is crossed with a semantic difference (between naturally occurring and man-made substances) applying to words within a circumscribed semantic field (foodstuffs). We show, first, that grammaticality judgments of a patient with semantic dementia are indistinguishable from those of a group of age-matched controls, and are similar regardless of the status of his semantic knowledge about the item. In a second experiment we use the triadic task in a group of age-matched controls to show that similarity judgments are influenced not only by meaning (natural vs. manmade), but also implicitly by syntactic information (count vs. mass). Using the same task in a patient with semantic dementia we show that the semantic influences on the syntactic dimension are unlikely to account for this pattern in normals. These data are discussed in relation to modular vs. nonmodular models of language processing, and in particular to the semantic-syntactic distinction.
Introduction Semantics and syntax constitute two fundamental properties of natural languages. At the level of single words, meaning is embodied in what appears to be a somewhat arbitrary mapping between sound and referent (there is nothing canine about the word “dog”), while syntax defines the conventions according to which words are used in sentences. This distinction raises the issue, fundamental to the cognitive neuropsychology of language, of whether semantics and syntax are processed in the brain by functionally independent systems, or rather reflect different constraints on the operation of a single system. An extreme version of the former view has maintained that a system dedicated to the representation and processing of syntactic information is innate and even genetically determined (Pinker, 1994), while other formulations merely posit the functional separation of two distinct types of processing. In contrast to this modular view, however, connectionist formulations have been proposed, in which syntax and semantics alike are regarded as sources of constraint on a common distributed system for language processing (Bates and MacWhinney, 1989; Bates et al. 1995). Among the earliest studies to suggest the separability of semantics and syntax were those which described contrasting patterns of neuropsychological deficit in patients with acquired
brain lesions. Such studies have generally been in agreement that appreciation of word meaning tends to be disrupted by lesions affecting an area of cortex in the posterior temporal region of the left hemisphere, while syntax may, in contrast, be vulnerable to smaller left hemisphere lesions, in the vicinity of Brodmann area 44. This distinction is perhaps most vividly illustrated by the contrast between Broca’s and Wernicke’s aphasia, which predictably follow lesions to anatomically distinct regions of the left hemisphere language areas, and which either disrupt word meaning to a much greater extent than sentence structure (Wernicke’s) or vice versa (Broca’s). Although the significance of this apparent double dissociation of acquired aphasia subtypes has been subject to considerable debate (for a concise review, see Breedin and Sarran, 1999), two other lines of evidence have more recently been adduced to argue for the separability of meaning and form processing: the first comes from the technique of functional brain imaging, and the second from the study of patients with neurodegenerative conditions. There has been general accord among functional imaging studies that tasks requiring the appreciation of single-word meaning activate a widely distributed left hemisphere “language network” incorportating both prefrontal and posterior
Correspondence to: Dr. Peter Garrard, Institute of Cognitive Neuroscience, Alexandra House, 17 Queen Square, London WC1N 3AR, email:
[email protected] DOI: 10.1080/13554790490892248
354
P. Garrard et al.
temporal regions (see Price, 1998 for a review), while processing of syntactically complex sentences activates smaller areas of the inferior frontal cortex (Dapretto and Bookheimer, 1999; Vigliocco, 2000; Hashimoto and Sakai, 2002). Direct comparison of the activations obtained during tasks requiring semantic and syntactic processing, on the other hand, reveals that both activate regions of the left inferior frontal lobe, and that areas activated by tasks requiring semantic processing are frequently anterior to those activated by syntactic tasks. One of the clearest demonstrations of this effect came from a study conducted by Dapretto and Bookheimer (1999), in which participants were asked to judge whether pairs of sentences had the same or different meanings. Sentences with similar meanings required the subjects to compare two words (as in “the car is in the garage” vs. “the auto is in the garage”) or two syntactic forms (e.g., “the policeman arrested the thief” vs. “the thief was arrested by the policeman”). Subtraction of the activation patterns obtained during these two comparison conditions revealed that the syntactic task was associated with activation in BA 44 (Broca’s area), while the semantic task activated a rather more anterior portion of the left inferior frontal gyrus (BA 47). Studies of progressive neurodegenerative conditions have been rather fewer in number, reflecting the relative rarity of patients with appropriately circumscribed deficits. Several case descriptions of individuals or groups with semantic impairment have remarked on the apparent preservation of syntactic processes (Hodges and Patterson, 1996; Hodges et al., 1998) and two detailed single case studies have systematically demonstrated that preserved syntactic processing can take place in the face of profound semantic loss (Breedin et al., 1994; Breedin and Saffran, 1999), thus providing strong support for the modular nature of the language processor. In contrast, however, a group study of patients with Alzheimer’s disease (AD) demonstrated a parallel decline in grammatical complexity and word retrieval in spontaneous utterances, a finding that was taken to support the nonmodular view (Bates et al.,1995). One potential problem with both the functional imaging and patient studies reviewed above is that, while performance on a semantic task engaged the processing of single words, syntactic tasks required the processing of whole sentences. This raises the possibility that cognitive resources, such as attention and working memory, may be differentially recruited depending on whether the task involved a semantic or syntactic judgment. The importance of this for drawing conclusions from even the most carefully designed functional imaging studies, is that the critical comparison may as a consequence reflect differences in strategic processes rather than in the differential recruitment of specialist systems. A similar caveat must also apply to conclusions drawn from individuals or populations with semantic breakdown. One way of overcoming such difficulties is to limit the scope of investigation to the level of single words. A large number of researchers working in psycholinguistics argue that at least certain types of syntactic properties are lexically
represented (Levelt, 1989; McDonald et al., 1994; Vigliocco and Hartsuiker, 2002). According to this view, lexical representations incorporate both semantic and syntactic information: the former corresponds to the intended meaning for the lexical entry to become activated, while the latter constrains the sentence contexts and/or grammatical forms in which it is licensed to appear. Given these assumptions, it should in principle be possible to demonstrate dissociation at the single word level between functionally independent systems—if they exist—that are specialized for the processing of lexical semantic and syntactic information. Attempts to achieve this in the past by pointing to dissociations between words belonging to different grammatical categories (e.g., nouns and verbs), in patients (McCarthy and Warrington, 1985; Zingeser and Berndt, 1990; Bak et al., 2001) and functional neuroimaging experiments (Warburton et al., 1996; Tyler et al., 2001) have not yielded entirely consistent findings. One reason for this might be the close relationship that exists between grammatical class and semantics (nouns typically refer to objects and verbs to actions)— a confound that causes difficulties even for studies in which a noun-verb distinction is evident between homonymous pairs (e.g., the watch vs. to watch). The approach we take in the present study overcomes these difficulties by using two key modifications of previous techniques to compare semantics and syntax. The first is to combine both types of processing within a single task by systematically varying the semantic and syntactic content of items within a single stimulus set. The second is to use as the syntactic manipulation not the distinction between nouns and verbs, but that between count nouns (i.e., nouns that can be pluralized, and used with the determiner “a”) and mass nouns (nouns that cannot be pluralized and are used with the quantifier “some”). In order to ensure that there were no differences at the level of processing resources between the semantic and syntactic conditions, we employed a methodology that relies on the subject making semantic judgments on single words, namely, the triadic judgment task. In this task, participants are presented with triplets of words, asked to judge which two are most similar in meaning, and to respond by crossing the odd one out. This methodology has a long history in investigations of the representation of semantic knowledge, and its early disruption in patients with AD (Chan et al., 1995; 2001a) The method has been shown to be sensitive not only to semantic but also to syntactic similarity for a subset of verbs, even if the task does not explicitly require participants to use the syntactic information (Fisher, 1994). We hypothesized that: 1) performance on the triadic task by normals would reflect semantic similarities between words as well as similarities on the syntactic (count vs. mass) dimension, and 2) that in accordance with the general neuropsychological profile of such patients, sensitivity to the semantic, but not the syntactic aspects of individual words would be disrupted in a patient with semantic impairment. We referred earlier to the semantic confound that has bedevilled attempts to use nouns and verbs to represent a
Dissociation of lexical syntax and semantics
word-level syntactic distinction, and anticipate that a similar objection may be raised here. We certainly acknowledge that the count/mass distinction is likely to have a semantic foundation, but would argue that it is not completely reducible to semantics, on a number of grounds. First, the differentiation of certain count/mass pairs, such as noodles/spaghetti or opinion/ knowledge, has no obvious explanation in semantics. Secondly, count/mass status can differ across languages: for instance, the word for spinach is mass in English, but count in Italian; in Hebrew the word for water is a count noun; and in Japanese the count/mass distinction does not exist. Thirdly, Vigliocco et al. (1999) showed that normal subjects who reported a ‘tip-of-the-tongue’ state during a naming to verbal description task could successfully guess the count/mass status of the target word only when they were experiencing a true lexical retrieval failure. Finally, Semenza et al. (1997) described an aphasic patient who had difficulty assigning quantifiers correctly to mass nouns, while performing normally on a range of other semantic and syntactic properties of the same words. We will argue that the data presented here—in particular the performance of a patient with semantic impairment—provide further support for a pure syntactic dimension to the count/ mass distinction, as well as suggesting that the processing of semantic and syntactic information may be functionally dissociable.
Subjects The studies were conducted on a patient with semantic impairment and on groups of cognitively normal participants, recruited through a volunteer database maintained at the Institute of Cognitive Neuroscience. Controls were matched as closely as possible to the patient for age and years of education. These demographic characteristics of the patient and controls are displayed in Table 1.
Patient ‘Oscar’ is a native English speaking male who left school at 16, spent several years in the armed forces, and then worked in a middle-ranking managerial position until he was made redundant in his early 50s. After some years’ employment in a succession of more menial jobs, a combination of depression and a multitude of somatic symptoms prevented further regular work. Investigation of his general symptoms by a hospital physician revealed systemic hypertension only, and combination antihypertensive therapy was instituted, as well as a variety of symptomatic treatments. Table 1. Demographic characteristics of participants
Age Years of education
Oscar
Experiment 1: Controls (mean [SD])
Experiment 2: Controls (mean [SD])
61 11
64.9 [8.2] 14 [3.2]
65 [6.9] 12.8 [2.9]
355
The patient was referred to the National Hospital for Neurology and Neurosurgery in 2000 because of early changes in behavior and personality noticed by his wife, including rigidity of routine and an increasing obsession with his physical symptoms, about which he would produce regular written reports. Later he began to develop signs of social disinhibition (such as talking to strangers and barking at his own dog in public), and an obsession with removing all his body hair. Recognition by Oscar’s wife of the significance of these early symptoms owed much to her knowledge of a strong history of young-onset dementia among her husband’s close relatives. Post mortem examination of the brain of one of his siblings had recently revealed severe tau pathology in the frontotemporal cortex and deep grey-matter. A splice site mutation at exon 10+16 of the microtubule-associated protein tau (MAPT) gene had been demonstrated in another symptomatic member of the same generation. When he was first examined at the National Hospital in January 2002 his Folstein MMSE was 25/30, and it was noted that his behavior was marginally inappropriate, that proverb interpretation was rather concrete and cognitive estimates poor. The most prominent deficits were found, however, in the domain of language: spontaneous speech was fluent and grammatically correct but empty of content, and he performed poorly on bedside tests of verbal fluency, picture naming and naming to verbal description. His definitions of unnamed items suggested that his nominal difficulties derived from degraded semantic knowledge: for instance, when shown a picture of a rhinoceros, he said: . . . lives in the desert, not in UK. Wild because of the shape. Fair size like a cow. Not friendly. Doesn’t bite. Similarly, in response to the word ‘tortoise’ he responded: Big animal, foreign. Warm countries. In a zoo. Almost a beast. Quite dangerous. Runs fast. Bulky build, more than a horse. In contrast, he displayed good knowledge of recent personal and public events, and normal visuoperceptual and calculation skills. The results of a preliminary neuropsychological examination are presented in Table 2. A magnetic resonance (MR) brain scan, performed in April 2002, revealed periventricular white-matter hyperintensities consistent with small-vessel disease, together with mild global
Table 2. Preliminary neuropsychological examinations Test
Oscar
Normal range/ percentile score
MMSE VIQ PIQ RMT faces RMT words GNT Verbal fluency S F Animals Trails A Weigl sorting
25 99 105 29 22 0
>26 >5th percentile >5th percentile >1st percentile
5 6 10 50th-70th percentile PASS
mean14.4; sd 4.6 mean13.7; sd 4.5 >10th percentile -
356
P. Garrard et al. Table 3. Results from serial assessments of Oscar’s performance on a series of tests of semantic and syntactic knowledge Test
Max. score
Oscar April Nov 02
Fig. 1. Representative coronal slices from T1 weighted volumetric MRI of subject Oscar, acquired May 2003, demonstrating profound bilateral temporal neocortical atrophy, particularly in the left hemisphere (right side of the scans), with relative sparing in frontal and parietal regions.
atrophy, disproportionately prominent in both temporal lobes, and relative preservation of the hippocampi, particularly in more posterior slices – in keeping with the clinical diagnosis of semantic dementia (SD) (Galton et al., 2001; Chan et al., 2001b). T1 weighted images acquired approximately one year after Oscar was first seen are reproduced in Fig.1.
Standardized assessments of semantic and syntax The relative disruption of Oscar’s semantic and syntactic processing was explored first using a range of standardised assessments, and later by a series of experimental studies. We sought, first, to confirm that Oscar was impaired at the level of semantic knowledge by administering a battery of semantic tests. This battery (described in detail elsewhere (Bozeat et al., 2000)) consists of a series of tasks requiring access to and from a common set of 64 items represented in both verbal and visual form, through a variety of channels (e.g., picture naming, naming to verbal definition, word-picture matching and associative matching). His pattern of performance on these tests at two testing rounds approximately 12 months apart is shown in Table 3, and indicates consistent difficulties regardless of the modality of presentation or response. Several further features of his test performance are worthy of note: First, Oscar shows a category-specific naming and comprehension deficit—an unusual but recognized pattern in the context of semantic dementia (Lambon Ralph et al., 2003). Secondly, deterioration in performance between testing rounds is limited to naming, indicating no more than a mild degree of progression over this period. Thirdly, his ability to understand similarities among concepts is better at superordinate (Picture sorting, levels 1 and 2) than more fine-grained levels— a pattern characteristic of semantic impairment. Finally, the greater difficulty with producing names in response to spoken definitions consisting of the item’s sensory features (e.g., “What
1) Semantic battery Picture naming living nonliving Naming to description sensory features functional features Word-to-picture matching Living Nonliving 2) Non-verbal semantic tests Camel and Cactus (pictures) Picture sorting level 1 (living/nonliving) level 2 (category) level 3 (specific feature) Sound to picture matching 3) Syntactic tests Grammaticality judgement TROG
Control mean [SD] MayJuly 03
64 32 32 128 64 64 64 32 32
36 13 23 52 35 17 45 14 31
32 10 22
64
34
36
56.9 [7.27]*
64 64 144 48
63 62 118 38
63 61 115 -
41.2 [2.5]*
80 80
78 78
49 20 29
76
62.3 [1.6]*
57.1 [2.4]† 54.0 [2.0]† 63.7 [0.5]*
70.8 [5.8] >75th percentile
*From Bozeat et al. (2000). †From Garrard et al. (2002).
do we call the large four-legged animal with hooves, and one or two humps on its back?”) than with those emphasizing functional features (e.g., “What do we call the bad-tempered animal that is ridden in the desert, and can survive without water?”), is a frequent finding in patients with semantic dementia (Garrard et al., 2002; Lambon Ralph et al. 2003). Also shown in Table 3 is Oscar’s performance on two tests of syntactic ability: a grammatical/nongrammatical judgement test on 80 sentences, of which half were grammatically wellformed and the remainder contained a range of grammatical violations, and the test for the reception of grammar (TROG) (Bishop, 1989). On both of these tests Oscar’s performance fell within the normal range. We went on to conduct two experimental studies designed to assess the patient’s performance on tests of the semantic and syntactic properties of single words. In the first, we compare his ability to appreciate the acceptability of a series sentences in which count and mass nouns are preceded by typical and atypical determiners, and look for any relationship between his ability to make such judgments and his semantic knowledge of the same item. In Experiment 2 we introduce the triadic task, and report the results from controls (2a) and from the patient (2b).
Experiment 1 Materials and Methods In addition to the 64 items in the semantic battery detailed above, Oscar had been presented with a range of other items in confrontation naming and spoken word-to-picture matching
Dissociation of lexical syntax and semantics
Results Controls provided a range of ratings for each sentence, broadly reflecting the most commonly accepted count or mass reading of the items. When mean control ratings for each item were compared in a 2 (count vs. mass) by 2 (typical vs. atypical construction) analysis of variance, the typical sentences received significantly higher ratings than the atypical [F(1,44) = 69.35, p
