Introduction
Language impairments in schizophrenia (SZ), typically seen in the form of incoherence and poverty in speech, have been attributed to abnormalities of the semantic system and related cognitive functions, such as executive control (Allen, Liddle, & Frith, Reference Allen, Liddle and Frith1993; Gourovitch, Goldberg, & Weinberger, Reference Gourovitch, Goldberg and Weinberger1996; Leeson, Laws, & McKenna, Reference Leeson, Laws and McKenna2006; Rossell & David, Reference Rossell and David2006). One neuropsychological approach used to identify this impairment involves category-cued verbal fluency tasks. These typically require examinees to report as many exemplars of a semantic category, such as animals, as possible in a limited time, often 60 seconds. Persons with SZ generate fewer words than healthy adults, with an average effect size of 1.2 standard deviations (Bokat & Goldberg, Reference Bokat and Goldberg2003). They also produce smaller clusters of semantically related words and show inefficient switching among clusters (Moelter et al., Reference Moelter, Hill, Ragland, Lunardelli, Gur, Gur and Moberg2001; Robert et al., Reference Robert, Lafont, Medecin, Berthet, Thauby, Baudu and Darcourt1998; Zakzanis, Troyer, Rich, & Heinrichs, Reference Zakzanis, Troyer, Rich and Heinrichs2000).
Another approach to detecting semantic system abnormalities in SZ is to construct a visual representation of the semantic clustering of words reported on fluency tasks. Since exemplars of a given category are often reported in subgroup clusters, such as pets, birds, and fishes on animal naming (Troyer, Moscovitch, & Winocur, Reference Troyer, Moscovitch and Winocur1997), clustering patterns might provide clues to the nature of semantic system abnormalities in SZ. Multi-dimensional scaling (MDS) is often used for this purpose. MDS studies have shown that patients with SZ show less coherent clustering of the same exemplars than healthy controls (Aloia, Gourovitch, Weinberger, & Goldberg, Reference Aloia, Gourovitch, Weinberger and Goldberg1996; Moelter et al., Reference Moelter, Hill, Ragland, Lunardelli, Gur, Gur and Moberg2001; Paulsen et al., Reference Paulsen, Romero, Chan, Davis, Heaton and Jeste1996; Rossell, Rabe-Hesketh, Shapleske, & David, Reference Rossell, Rabe-Hesketh, Shapleske and David1999; Sumiyoshi et al., Reference Sumiyoshi, Sumiyoshi, Nohara, Yamashita, Matsui, Kurachi and Niwa2005). One possible explanation of this is that persons with SZ have degraded/disorganized semantic networks (Aloia et al., Reference Aloia, Gourovitch, Weinberger and Goldberg1996; Bozikas, Kosmidis, & Karavatos, Reference Bozikas, Kosmidis and Karavatos2005; Laws, Al-Uzri, & Mortimer, Reference Laws, Al-Uzri and Mortimer2000; Paulsen et al., Reference Paulsen, Romero, Chan, Davis, Heaton and Jeste1996; Rossell & David, Reference Rossell and David2006; Rossell et al., Reference Rossell, Rabe-Hesketh, Shapleske and David1999). Another is that SZ disrupts access to or retrieval of elements from the semantic network (Allen & Frith, Reference Allen and Frith1983; Allen et al., Reference Allen, Liddle and Frith1993; Joyce, Collinson, & Crichton, Reference Joyce, Collinson and Crichton1996). Such a functional impairment could result from disrupted associations among semantic elements.
Unfortunately, MDS analysis cannot determine the source of semantic system abnormalities. Differences in visually represented semantic clusters given by persons with SZ and NCs do not necessarily signify degraded semantic elements. They might denote functional deficits of the semantic system and cognitive control (Storms, Dirikx, Saerens, Verstraeten, & De Deyn, Reference Storms, Dirikx, Saerens, Verstraeten and De Deyn2003), such as the automatic activation or inhibition of related concepts within a semantic cluster or strategic searching for subcategories (e.g., imagining supermarket aisles). Both types of deficits are also possible since they may not be mutually exclusive. Practical difficulties also constrain the application of MDS to category fluency data. As a result, the interpretation of MDS findings is limited to general statements about differences in the semantic clustering patterns of persons with SZ.
The main goal of the current study is to investigate and compare the details of semantic organization underlying category fluency productions by patients with SZ and healthy adults. We adapted a matrix factorization procedure called singular value decomposition (SVD) for this purpose. SVD is a general matrix factorization technique of which eigenvalue decomposition, the mathematical basis for factor analysis, is a special case. This method has been used in many areas of science (Alter, Brown, & Botstein, Reference Alter, Brown and Botstein2000; Landauer, Reference Landauer2007), but it has not been applied to verbal fluency data. Using SVD, we aimed to identify some critical differences in semantic knowledge organization and its functions between patients with SZ and healthy controls, differences that have not been reported previously based on different clustering analyses. We will first describe limitations of using MDS with fluency data to elucidate semantic system abnormalities in SZ and explain the potential advantages of using SVD for this purpose. (Details of how we applied SVD to category fluency are provided in Supplementary Information.)
Limitations of MDS Studies of Category Fluency
Typically, MDS requires proximity measures that reflect the strength of association between all possible pairs of words to be analyzed. These distance measures are not directly available from category fluency data. Word sequence has been used to construct a proximity matrix based on the assumption that semantically related words tend to be reported closely in sequence (Chan, Butters, Salmon, & Mcguire, Reference Chan, Butters, Salmon and Mcguire1993).Footnote 1 For example, in the sequence: cow—pig—sheep—lion—giraffe—elephant, “cow” is thought to be more semantically related to “sheep” than to “elephant” because the distance between cow and sheep is two and the distance between cow and elephant is five. However, calculating a proximity matrix in this way can yield estimates that contradict common sense. In this same sequence, the word order suggests that “sheep” and “lion” are more closely related than “cow” and “sheep.” This obviously is wrong. More likely, the sequence includes two subcategories: farm animals and African animals (Troyer et al., Reference Troyer, Moscovitch and Winocur1997). From this perspective, the within-subcategory exemplars “sheep” and “cow” are more semantically related than the between-subcategory exemplars “sheep” and “lion,” even though the latter were reported closer together in sequence. Thus, proximity measures based on word order can distort visual representations of the semantic network. A counter argument is that the average group statistics should eventually give us proximity matrices that reflect the subcategory grouping of words.
A more critical limitation is that MDS studies never use more than the 20 or so most commonly reported exemplars of a category to construct proximity matrices (Aloia et al., Reference Aloia, Gourovitch, Weinberger and Goldberg1996; Chan, Butters, Paulsen, et al., Reference Chan, Butters, Paulsen, Salmon, Swenson and Maloney1993; Chan & Ho, Reference Chan and Ho2003; Paulsen et al., Reference Paulsen, Romero, Chan, Davis, Heaton and Jeste1996; Rossell et al., Reference Rossell, Rabe-Hesketh, Shapleske and David1999; Sumiyoshi et al., Reference Sumiyoshi, Sumiyoshi, Nohara, Yamashita, Matsui, Kurachi and Niwa2005). The reason is that using lower frequency words leads to missing values in many cells of the proximity matrix. This is a known limitation of using MDS with verbal fluency data (Giovannetti, Goldstein, Schullery, Barr, & Bilder, Reference Giovannetti, Goldstein, Schullery, Barr and Bilder2003; Rossell et al., Reference Rossell, Rabe-Hesketh, Shapleske and David1999), one that could bias the results of MDS. Furthermore, most MDS studies report only two-dimensional solutions based on the interpretability of observed dimensions and goodness-of-fit measures. But the underlying structure might be better represented by three or more dimensions if one could construct a larger proximity matrix with few missing values. Under these conditions, MDS might reveal more differences between patients and controls. Two dimensions found by several MDS studies of animal naming involve domesticity and size (Aloia et al., Reference Aloia, Gourovitch, Weinberger and Goldberg1996; Chan, Butters, Paulsen, et al., Reference Chan, Butters, Paulsen, Salmon, Swenson and Maloney1993; Chan & Ho, Reference Chan and Ho2003; Paulsen et al., Reference Paulsen, Romero, Chan, Davis, Heaton and Jeste1996; Rossell et al., Reference Rossell, Rabe-Hesketh, Shapleske and David1999). However, the size dimension seems to emerge only from some other MDS and non-MDS studies, but not from others (Giovannetti et al., Reference Giovannetti, Goldstein, Schullery, Barr and Bilder2003; Moelter et al., Reference Moelter, Hill, Ragland, Lunardelli, Gur, Gur and Moberg2001; Troyer et al., Reference Troyer, Moscovitch and Winocur1997). Thus, some dimensions found by MDS might be artifacts of word input constraints.
These limitations do not suggest that including more words in an MDS analysis would necessarily yield a better representation of the semantic structure. They do suggest that it is bold to assume that hidden semantic structures can be elucidated by analyzing just 20 words. Basing an MDS analysis on 20 words also is likely to obscure critical differences between the semantic systems of patients and controls. It is unlikely that MDS can overcome the small word input problem in category fluency analyses (Storms et al., Reference Storms, Dirikx, Saerens, Verstraeten and De Deyn2003). We believe that SVD offers the flexibility required to overcome this limitation and better elucidate abnormalities of semantic clustering shown by patients with SZ.
Method
Participants
Patients were recruited for two studies of SZ from Johns Hopkins-affiliated outpatient clinics, inpatient units, a day hospital, and via flyers posted locally. They included 102 adults who met DSM-IV criteria for schizophrenia (American Psychiatric Association, 1994). Diagnoses were made by a study psychiatrist using the Diagnostic Interview for Genetic Studies (Nurnberger et al., Reference Nurnberger, Blehar, Kaufmann, York-Cooler, Simpson, Harkavy-Friedman and Reich1994). Participants with any history of intellectual disability, dementia, stroke, traumatic brain injury, or substance dependence (within the past 12 months) were excluded. Our sample included 102 treated adults with SZ. Of these, 24 were classified as having the deficit syndrome, 76 were classified as having non-deficit SZ, and 2 were not classified. Most (93%) had at least one prior psychiatric hospitalization (M = 5.3, SD = 5.8). The mean age of diagnosis was 23.0 (SD = 7.5) years. Of the 90 patients for whom information about current medication use was deemed reliable, the medications included taking typical neuroleptics (33%), atypical neuroleptics (64%), both typical and atypical neuroleptics (13%), antidepressant (24%), anticonvulsant (12%), and lithium (4%). Most patients were assessed several years (M = 6.1; SD = 8.9) years after their most recent hospitalization. Eight were examined immediately before being discharged from the hospital when their attending physicians determined they were stable. The patients generally were rated as showing minimal to moderate symptoms based on Andreasen and Olsen's (Reference Andreasen and Olsen1982) Scales for the Assessment of Negative and Positive Symptoms (SANS M = 8.7; SD = 5.4 and SAPS M = 4.7; SD = 3.7, respectively). Demographic characteristics and estimated pre-morbid IQ based on the Hopkins Adult Reading Test (HART; Schretlen et al., Reference Schretlen, Winicki, Meyer, Testa, Pearlson and Gordon2009) of the patients are shown in Table 1.
Table 1 Demographics of NC and SZ groups
1Based on Hopkins Adult Reading Test (HART; Schretlen et al., Reference Schretlen, Winicki, Meyer, Testa, Pearlson and Gordon2009).
Note. * p < .05.
The patients were matched to a healthy control (NC) group drawn from the Aging, Brain Imaging, and Cognition (ABC) study of normal aging in a community sample (Schretlen, Testa, Winicki, Pearlson, & Gordon, Reference Schretlen, Testa, Winicki, Pearlson and Gordon2008). The NC group was recruited via random digit dialing or calling randomly selected listings from residential telephone directories in the Baltimore, Maryland and Hartford, Connecticut metropolitan areas. In addition to the exclusion criteria listed for SZ participants, ABC study participants were excluded from analysis if they had a current psychiatric, medical, or neurological condition associated with cognitive impairment, or if they scored below 24/30 on the Mini-Mental State Exam (Folstein, Folstein, & McHugh, Reference Folstein, Folstein and McHugh1975). This resulted in usable data for 327 NCs from which we selected 109 participants who matched the SZ sample in sex, age, race, and estimated pre-morbid IQ. As shown in Table 1, the patients with SZ completed significantly fewer years of education that the healthy adults. To avoid the “matching fallacy” (Meehl, Reference Meehl1970) of selecting under-educated NCs by matching their educational background to that of the patients with SZ, we matched the groups on estimated premorbid IQ. In this way, we attempted to minimize the effects of group differences in educational level and maximize the homogeneity of pre-morbid IQ, an approach we have used previously (Schretlen et al., Reference Schretlen, Cascella, Meyer, Kingery, Testa, Munro and Pearlson2007). The Johns Hopkins Medicine Institutional Review Board approved the studies from which subjects were drawn, and each person gave written informed consent to participate.
Procedure
Each participant completed two category-cued (animals and supermarket items) oral word fluency tasks taken from the Calibrated Ideational Fluency Assessment (CIFA; Schretlen & Vannorsdall, Reference Schretlen and Vannorsdall2010) as part of a larger neurocognitive assessment. On each task, participants were asked to say as many category examples as they could think of in 60 s. Their responses were recorded verbatim by the examiner.
Analysis
For analysis, we removed rule breaks, such as repeated words and non-category examples, from the word lists. There were 340 rule breaks among the 9952 animal words (3.4%) given by all participants. Patients made fewer rule breaks than healthy controls (102 vs. 238). There were 715 rule breaks out of total 12,847 supermarket words (5.6%) reported by all participants, and 282 of these were given by patients. Plurals were converted into singular form in the animal condition and singulars to plural form in the supermarket item condition for consistency in grammatical number. A repeated measure analysis of variance (ANOVA) with mixed design was then performed on the number of correct exemplars produced with subject group (between-group) and the category (within-group) as factors.
One matrix per category cue condition was constructed for both groups for SVD analysis (four matrices in total). Each matrix consisted of rows of different words and columns of subjects. The 109 NCs reported 251 unique animals and 486 unique supermarket items, yielding 251 × 109 (animals) and 486 × 109 (supermarket items) matrices for SVD analysis. The 102 patients reported 215 unique animals and 423 unique supermarket items, yielding input matrices of 215 × 102 and 423 × 102, respectively. Each cell (cij) of the matrices was assigned a value of 1 if person j gave the word i and 0 otherwise. We first sought 25-dimensional SVD solutions for all four matrices using PROPACK software (Larsen, Reference Larsen2004) for Matlab (Version 7.8, Mathworks). While arbitrarily chosen, we assumed that the number of meaningful dimensions would be smaller than 25. Since our goal is to identify critical differences of semantic clustering patterns between the two groups, we did not try to find the specific dimensional solutions (i.e., specific dimensionality) that fit the data best. Furthermore, we present only the 40 most frequently reported words for each category by the combined samples. Although all possible category examples were analyzed, examples with frequency ranks below 40 were not very informative. (See the Supplementary Information for full dimensional SVD solutions.)
Results
Statistical Analysis
As shown in Table 2, the SZ group reported fewer words than the NCs on both CIFA subtests. Both groups reported more supermarket items than animal names. A repeated measure ANOVA confirmed this: There were significant main effects of participant group and task; F(1,209) = 46.72, p < .001, partial eta squared (ηp2) = 0.18, and F(1,209) = 113.35, p < .001, ηp2 = 0.35, respectively. But there was no interaction between participant group and category fluency task, F(1,209) = 2.38, p = .13, suggesting that the difference between NC and SZ groups was approximately the same for both category fluency tasks.
Table 2 Descriptive statistics for category examples given by participants in the NC and SZ groups
SVD Analysis
Animal category: cluster space representations
Figure 1 shows the 40 most frequently reported animals by group, as 2-D representations between dimensions 2 and 3 (although the animal clusters are more clearly seen in a higher dimensional space, 2-D representations parsimoniously depict group differences). Other dimensional representations are provided in the Supplementary Information, along with greater detail about SVD methods. To better appreciate clustering, animals with low ranks (21–40) are shown in gray in Figure 1. The frequency ranks of words shown in all figures were based on their representation in a large category fluency database derived from healthy adults and diverse patient groups (described in Supplementary Information). Clusters are defined by the angles between vectors (cosines of angles, precisely) representing different animals. For example, consider monkey, hippopotamus, and elephant in SZ (Figure 1b). Hippopotamus co-occurs frequently with monkey (15 times in SZ). However, this is low relative to the absolute frequency of monkey (n = 54). This results in similar directions of vectors, with a shorter one for hippopotamus in the 2-D plot (it may be longer in other dimensions). Elephant (n = 54) co-occurs more frequently with monkey (31 times in SZ), but their vector angle is wider because elephant also co-occurs with other animals that are not associated with monkey. (See Supplementary Information for further details.)
Fig. 1 Forty most frequently and commonly named animals (ranks from 1 to 40) plotted in 2-D vector space (dim. 2 vs. 3). (a) Clustering of 40 animals by healthy adults. (b) Clustering of the same 40 animals by patients with SZ.
Based on vector angles, Figure 1a for NC suggests that the top 40 animals comprise at least three clusters: domestic/farm animals in the first and fourth quadrants, wild/African animals in the some of the first and second quadrants, and sea animals/reptiles/birds/fish in the third quadrant.Footnote 2 When more than 2 dimensions are plotted, sea animals (whale, dolphin, shark, and fish) and reptiles (crocodile, alligator, snake, lizard, and possibly bird) form separate clusters that cannot be seen in Figure 1a. Animals do not seem to cluster based on size. Another finding in the NC group is that word vectors representing the most frequently named animals (dog, cat, lion, and tiger) are not clearly subdivided into domestic (i.e., dog and cat) and wild/African (i.e., lion and tiger) clusters. As shown in Figure 1b, patients with SZ clustered wild/African animals and farm/domestic animals, but their lower ranked animals did not form any discernible clusters. Thus, unlike healthy adults, patients with SZ showed no clear cluster of shark, eagle, dolphin, lizard, whale, and other reptiles. Finally, the top ranked animals (cat, dog, lion, tiger) clearly sub-divide into relevant clusters in SZ, which differs from their clustering pattern in NCs.
Animal category: cosine measures of clustering
We calculated cosines of angles of three selected animal vectors (cat, cow, and whale) against the top 40 animal vectors and plotted them in Figure 2. These animals were chosen because they show some representative differences and similarities between the NC and SZ groups. The cosine values were calculated for first 2, 3, 4, and 5 dimensions, represented as dotted dark, solid dark, solid gray, and dotted gray lines in Figure 2. The clustering patterns revealed by cosine values seem most interpretable when the first three (1, 2, and 3) or four (1, 2, 3, and 4) dimensions are considered (solid dark and gray lines) both in animal and supermarket item conditions (see below). The x-axes of these plots represent the 40 top-ranked animals whose names are listed in the caption. To better appreciate these plots, animal names will be presented with their ranks in parenthesis in the supplementary section.
Fig. 2 (a–f) Word-word co-occurrence measured by cosines of angles of word vectors. In these plots, a vector for three pre-selected animal vectors (cat, cow, and whale) is paired with the top 40 animal vectors for cosine calculations in NC and SZ. Four different lines within each panel indicate different number of dimensions included for calculation of cosines. The numbers of x-axis indicate top 40 animals. Numbering by five, they are: cat (1), dog, lion, tiger, elephant (5), giraffe, bear, horse, zebra, monkey (10), snake, cow, bird, pig, deer (15), fish, mouse, rabbit, hippopotamus, rhinoceros (20), rat, alligator, squirrel, sheep, chicken (25), gorilla, whale, goat, leopard, eagle (30), crocodile, fox, kangaroo, shark, lizard (35), raccoon, ape, dolphin, duck, and donkey (40).
For NCs, cat (rank 1) has cosine values that reflect little clustering with other animals (Figure 2a). This also applies to other top-ranked animals [dog (2), lion (3), and tiger (4), not shown]. Although cat (1) tends to show higher cosine values (close to 1) with other domestic/farm animals for NC, this is far more evident in SZ (Figure 2b). That is, cosines in Figure 2b show more pronounced variations between cat (1) and the top 40 animals. For patients with SZ, cat (1) frequently co-occurs with dog (2) and other domestic/farm animals, but not with lion (3), tiger (4) and other wild/African animals. Among patients with SZ, the cosine between cat and dog is close to 1, whereas the cosines are closer to 0 for cat in relation to lion (3), tiger (4), elephant (5), giraffe (6), bear (7), and others. Other top ranked animals [dog (2), lion (3), tiger (4)] given by patients with SZ also show clustering patterns that are similar to those found for cat (1). These confirm the findings shown in Figure 1, namely; clear clustering of four top ranked animals among SZ but not NC.
When cow (12) is compared to the top 40 animals named by NCs (Figure 2c), the clustering pattern changes little regardless of how many dimensions are considered. This suggests that among NCs the concept of cow (12) is strongly associated with other domestic animals [e.g., horse (8), pig (14), rabbit (18), etc.] and distinct from animals that belong to other semantic clusters. In Figure 2d, patients show a similar clustering pattern for cow (12), except for dog (1) and cat (2) (for a reason explained above). The word cow (12) is more closely associated with dog (1) and cat (2) in SZ than in NC. Clustering patterns of the two groups diverge greatly from each other among lower ranked animals, with patients losing coherency of semantic clusters, as depicted in Figures 2e and f. Among lower ranked animals (ranks 21–40), whale (27) show the most representative differences between the NC and SZ groups.
In summary, the SVD results show a few large differences in clustering patterns for the 40 top ranked animals, based on the cosines of vector angles when more than 3-dimensional vector space is considered. However, large group differences were found for the four top ranked animals (dog, cat, lion, and tiger). The most striking differences between the SZ and NC groups were in their semantic associations to less frequently reported animals.
Supermarket category: cluster space representations
Figure 3 shows the top 40 supermarket items generated by both groups. In Figure 3a (NC), the fourth quadrant mainly includes dairy products. Specific fruits and vegetables cluster in the second quadrant. Between these two, specific meats (e.g., chicken, beef, turkey, etc.) appear to form a small cluster in the second and third quadrants, but they overlap examples of fruits and vegetables in this 2-D representation. Food subcategories (meat, vegetables, fruit, etc.) and non-food products (paper towels, toilet paper) form another cluster in the first quadrant.
Fig. 3 Forty most frequently and commonly named supermarket items (ranks from 1 to 40) plotted in 2-D vector space (dim. 2 vs. 3). (a) Clustering of 40 supermarket items by healthy adults. (b) Clustering of the same 40 supermarket items by patients with SZ.
Patients with SZ show very disorganized clustering patterns. In Figure 3b, it is nearly impossible to discern clusters that resemble those shown by NCs: Dairy products are mixed with vegetables; fruits and vegetables are separated; and the clustering patterns reflect little coherence, regardless of word frequency ranks, over the first three dimensions.
Supermarket category: cosine measures of clustering
The cosines of vector angles for three supermarket items are plotted in Figure 4. As before, the word ranks are in parentheses. As seen in Figure 3a, milk (1) tends to show larger cosines with other dairy products among NCs (Figure 4a) than patients with SZ (Figure 4b). Also, the cosine values by milk (1) tend to be more stable in the NC than SZ group, indicating strong semantic relationships among dairy product concepts in healthy adults.
Fig. 4 Word-word co-occurrence measured by cosines of angles of word vectors. In these plots, a vector for one of three selected supermarket items (cat, cow, and whale) is paired with the top 40 supermarket item vectors for cosine calculations in NC and SZ. Four different lines within each panel indicate different number of dimensions included for calculation of cosines. The numbers of x-axis indicate top 40 animals. Numbering by five, they are: milk (1), bread, cheese, eggs, apples (5), meat, chicken, lettuce, cereal, ice cream (10), oranges, soda, tomatoes, vegetables, potatoes (15), butter, fish, candy, bananas, fruit (20), carrots, cookies, cake, onions, steak (25), sugar, yogurt, soup, juice, pears (30), beef, toilet paper, ham, bacon, potato chips(35), coffee, celery, turkey, paper towels, and lunch meat (40).
When the word eggs (4) is compared to the 40 top ranked supermarket item exemplars (Figures 4c and 4d), the NC group shows comprehensible and stable clustering patterns in 2-, 3-, and 4-dimensional vector spaces. The clustering patterns shown by patients with SZ are quite different. Thus, in patients with SZ, eggs (4) clustered more closely with lettuce (8) and tomatoes (13) than with other breakfast foods, milk (1) and cereal (9). Another clear group difference is shown in Figures 4e and 4f: Among NC participants, lettuce (8) clustered with other fruits and vegetables, such as apples (5), oranges (11), tomatoes (13), but not with the subcategory names, vegetables (14) and fruits (20). This pattern is virtually invariant across four different dimensional solutions. In SZ, however, lettuce (8) is more associated with meat (6) than with apples (5) or oranges (11), although associations with some vegetables, such as tomatoes (13), potatoes (15), and carrots (21), are intact.
In summary, supermarket items generated by NC participants tend to cluster into dairy products, fruits/vegetables, non-food products and subcategory names (e.g., vegetables, fruits), and specific meats (bacon, chicken, and turkey). However, the patients with SZ showed little coherent clustering of the same supermarket items.
Discussion
The results of this SVD analysis confirm the finding of previous MDS studies that the category exemplars reported by persons with SZ form less coherent semantic clusters than exemplars reported by healthy adults. However, our results extend this general conclusion by showing more detailed, previously unreported differences in the semantic associations of persons with SZ. They also point to a possible source of the semantic system abnormality in SZ that cannot be shown by MDS.
In animal naming, 2-D clustering patterns and cosines of word vectors showed that patients with SZ possess comprehensible clusters of high-frequency animals (rank 1–10 or so). In fact, SZ patients showed stronger clustering of the four most frequently reported animals (dog, cat, lion, and tiger) than NCs (Figure 1). Conversely, they showed virtually no identifiable clustering of lower ranked animals. Healthy adults showed clear evidence of semantic organization much “deeper” into their animal naming output. Animal names by healthy adults were roughly organized into clusters of farm/domestic animals, wild/African animals and sea animals/reptiles. A few MDS studies (Paulsen et al., Reference Paulsen, Romero, Chan, Davis, Heaton and Jeste1996; Sumiyoshi et al., Reference Sumiyoshi, Sumiyoshi, Nohara, Yamashita, Matsui, Kurachi and Niwa2005) found the first two—but not the last—of these clusters, probably because MDS studies have been unable to analyze animal names in the third cluster. Contrary to many previous MDS studies (Aloia et al., Reference Aloia, Gourovitch, Weinberger and Goldberg1996; Chan, Butters, Paulsen, et al., Reference Chan, Butters, Paulsen, Salmon, Swenson and Maloney1993; Chan, Butters, Salmon, et al., Reference Chan, Butters, Salmon and Mcguire1993; Paulsen et al., Reference Paulsen, Romero, Chan, Davis, Heaton and Jeste1996; Rossell et al., Reference Rossell, Rabe-Hesketh, Shapleske and David1999), we found no animal clusters based solely on size. We suspect that the size dimension is an artifact of having too few category exemplars for MDS analysis. Our results further show that the dimensionality of vector space, at least in NCs, could reach a fourth dimension of associative organization.
On the supermarket fluency task, differences between the semantic associative clustering of patients with SZ and healthy adults were even more striking. The patients showed negligible evidence of normal clustering for any supermarket items. Semantic clusters shown by healthy adults include dairy products, specific fruits and vegetables, food subcategories/non-food items, and specific meats.
Possible Source of Deficit in Semantic System of Patients With SZ
Two different types of semantic system abnormality could explain impoverished category word fluency in SZ. One possibility is that their semantic knowledge is impoverished or structurally degraded (Bozikas et al., Reference Bozikas, Kosmidis and Karavatos2005; Laws et al., Reference Laws, Al-Uzri and Mortimer2000; Paulsen et al., Reference Paulsen, Romero, Chan, Davis, Heaton and Jeste1996; Rossell & David, Reference Rossell and David2006). This can be conceptualized as a structural deficit. Another is that they suffer from a functional impairment of retrieval from (or access to) intact semantic knowledge (Allen & Frith, Reference Allen and Frith1983; Allen et al., Reference Allen, Liddle and Frith1993; Aloia et al., Reference Aloia, Gourovitch, Missar, Pickar, Weinberger and Goldberg1998; Joyce et al., Reference Joyce, Collinson and Crichton1996; Spitzer, Maier, & Weisbrod, Reference Spitzer, Maier and Weisbrod1997). Though not the intended purpose of SVD, two aspects of our results point to a functional deficit of retrieval/access as the core semantic fluency impairment in SZ.
First, if the impoverished verbal fluency shown by persons with SZ were due solely to degraded semantic knowledge, then the category exemplars they report should be those that are most intact. However, the patients showed relatively clear semantic clustering of animal names—but not supermarket items. Since high frequency exemplars probably are less vulnerable to knowledge degradation (Rossell & David, Reference Rossell and David2006), this disassociation in the clustering of high frequency examples across categories is difficult to explain based on a structural deficit in SZ. That is, if a loss of concepts is the core determinant of impaired semantic fluency in SZ, then the clustering of successfully named exemplars by patients with SZ should be comparable to the same words named by NCs since the context-guided retrieval process is supposed to be intact. Furthermore, the different coherence of successfully named high-frequency exemplars in two category conditions suggests that the effect of semantic context for word retrieval might differ between persons with SZ and healthy adults (e.g., less effective contextual guidance in naming supermarket items). Thus, it seems that an additional mechanism is required to explain the less coherent clustering of successfully named concepts. A functional impairment of lexical retrieval from the semantic store is a reasonable candidate. One might wonder if the smaller number of unique animals reported by patients (215) than controls (251) reflects some degeneration of semantic knowledge in SZ. However, while degeneration of semantic knowledge will reduce productivity, reporting fewer unique exemplars does not exclusively signify degraded semantic knowledge. Patients with SZ might report fewer animal names than healthy adults due to loss of memory traces (degeneration), impaired access to or retrieval of them, or both.
Second, the four most frequently reported animals were clustered by patients with SZ but not by NCs. While perhaps surprising, this cluster-independent pattern of high frequency examples by the NC group is not truly new. In an MDS study, Aloia et al. (Reference Aloia, Gourovitch, Weinberger and Goldberg1996) found two 2-D solutions of semantic space for animal names reported by NC and SZ groups. Their MDS solution for the NC group showed four top rank animals (dog, cat, lion, and tiger) located around the origin of the axes. This means that the Euclidean distances from each of these words to the others were similar, suggesting that these four exemplars showed the same cluster independence that we found. To make things more interesting, Aloia et al. (Reference Aloia, Gourovitch, Weinberger and Goldberg1996) also found that patients with SZ showed clearer clustering of these top ranked animals than NCs, just as we did. This failure of NCs to cluster the most frequently reported animals was also found in one other MDS study (Sumiyoshi et al., Reference Sumiyoshi, Sumiyoshi, Nohara, Yamashita, Matsui, Kurachi and Niwa2005). However, this intriguing difference has not been highlighted by other researchers, including Aloia et al. Since it emerged from two studies that used completely different clustering techniques than we did, it seems unlikely that this finding is due to chance.
One possible explanation of the NC participants’ seemingly counter-intuitive failure to cluster the most frequently reported animals is that it is a mathematical consequence of their high rate of co-occurrence. Thus, even though dog and cat belong to the same cluster in healthy adults, SVD might fail to show it because they co-occur with virtually every other reported animal. This possibility is supported by the co-occurrence frequency of these words: 94 (86%) of 109 healthy adults and 78 (76%) of 102 patients with SZ said both dog and cat. In contrast, whereas 81% of healthy adults reported both dog and lion, only 47% of patients with SZ did the same. Since dog and cat co-occur much more frequently than dog and lion in SZ, the vector for dog is close to the vector for cat but farther from the vectors for lion and tiger. In healthy adults, this does not happen because most NC participants reported all four of the top ranked animals together in their protocols. Similar patterns are observed with cat and tiger.
This pattern obviously does not mean that dog and cat (or lion and tiger) are semantically unrelated among healthy adults. When a healthy person reports dog and cat, one of the two words usually follows the other immediately. The probability that dog is reported at position t+1 given that cat is generated at position t within a protocol (i.e., PNC[dogt+1|catt]) is 0.32 (= 31/98) in NC. The same probability in SZ (PSZ[dogt+1|catt]) is 0.42 (= 35/83). Likewise, the PNC[catt+1|dogt] is 0.59 (=58/99) and PSZ[catt+1|dogt] is 0.44 (= 36/82). In contrast, the probability that lion is reported immediately after dog in NC (PNC[liont+1|dogt]) is 0.01 (= 1/98), even though lion and dog co-occur 88 times. Also, the PSZ[liont+1|dogt] is 0.0 (= 0/82). Thus, the conditional probabilities of reporting semantically associated top rank animals in succession are very high in both NC and SZ, whereas the conditional probabilities of these four top rank animals that are not semantically associated with each other (e.g., PNC[lionk+1|dogk]) are very low in both NC and SZ.
These co-occurrence frequencies and conditional probabilities suggest that healthy adults almost always generate dog and cat, followed by other domestic/farm animals. They then switch to lion, tiger and other wild/African animals, probably initiated by the same two animals as well. This suggests that associations between the cue “animal names” and the four top ranked exemplars (dog, cat, lion, and tiger) are extraordinarily strong. The strength of their association might even distinguish them from other animal concepts, which need more contextual guidance for retrieval. Patients with SZ also name these top ranked animals, and within the appropriate subcategories, but they are less likely to do so than healthy adults. The strengths of association of these exemplars to the cue “animal names” might be compromised enough for patients with SZ to require more contextual guidance (or mental effort) for retrieval, leading patients to cluster these exemplars like lower ranked animals.
Limitations and Conclusions of the Current Study
The current study has limitations. First, one might question the reliability of clustering patterns shown by healthy adults, since no previous study has reported such detailed information about them. Also, one might argue that the fact patients with SZ reported the same words named by NCs, but with lower frequency, might explain the group differences we observed. However, two randomly selected subgroups of 109 healthy adults showed very similar clusters of animals (see Supplementary Information), which were not found in a larger number of patients with SZ. This suggests that the clusters shown by healthy adults are fairly stable and that a simple difference in word frequency cannot explain the groups’ divergent patterns. Also, clustering of top ranked animals seems to be comparable to what is reported in previous studies using MDS, which partially validate our results. Second, we did not provide any objective goodness-of-fit measure of the SVD solutions because there is no well-established way to determine the optimal number of dimensions for SVD (Quesada, Reference Quesada2007). The number of dimensions for an SVD solution usually is determined by the researcher based on their interpretability. This is similar to factor analysis, where the optimal number of factors is often determined in the same ways (recall that factor analysis is a special case of SVD). Moreover, we did not pursue any goodness-of-fit measure because our goal was to identify new and more detailed differences between persons with SZ and healthy adults, rather than to find the best solution for healthy adults. We believe that we achieved this goal without looking for optimal semantic space solutions for healthy adults. Third, we did not conduct an MDS of the same data for comparison with SVD. Such a comparison would be exceedingly difficult because many technical and theoretical aspects would need to be addressed. However, we compared our SVD results with those of previously reported MDS analyses in the Discussion. The clustering pattern of high ranked animals found via SVD is similar to that found by some MDS studies (Aloia et al., Reference Aloia, Gourovitch, Weinberger and Goldberg1996; Sumiyoshi et al., Reference Sumiyoshi, Sumiyoshi, Nohara, Yamashita, Matsui, Kurachi and Niwa2005). We cannot compare lower-ranked animal clusters because low frequency category exemplars are precisely what MDS is not well-suited to analyze. In conclusion, we demonstrated more specific aspects of semantic system dysfunction in SZ on category word fluency tasks than recognized previously using SVD to overcome limitations of MDS. We conclude that the semantic deficit in SZ likely involves a functional impairment of access to or retrieval of semantic knowledge. Our findings do not exclude the possibility that persons with SZ have a structural degradation of semantic knowledge, but they are difficult—if not impossible—to explain solely on this basis.
Acknowledgments
This research was supported by The Therapeutic Cognitive Neuroscience Fund (BG), the Benjamin and Adith Miller Family Endowment on Aging, Alzheimer's, and Autism Research (BG), NIMH grants MH60504 and MH43775 (GDP and DJS), and the National Alliance for Research on Schizophrenia and Depression (DJS). Under an agreement with Psychological Assessment Resources, Inc., Drs. Schretlen and Vannorsdall are entitled to a share of royalties on sales of a test used in the study described in this article. The terms of this arrangement are being managed by the Johns Hopkins University in accordance with its conflict of interest policies.
Supplementary Materials
To review these additional data and analyses, please access the online-only supplementary text, table, and Figures 1-5. Please visit journals.cambridge.org/INS, then click on the link “Supplementary Materials” at this article.
