Artificial grammar learning in individuals with severe aphasia.

Neuropsychologia 53 (2014) 25–38

Contents lists available at ScienceDirect

Neuropsychologia journal homepage: www.elsevier.com/locate/neuropsychologia

Artificial grammar learning in individuals with severe aphasia Vitor C. Zimmerer a,n, Patricia E. Cowell b, Rosemary A. Varley a a b

Division of Psychology and Language Sciences, University College London, Chandler House, 2 Wakefield Street, London, WC1N 1PF, United Kingdom Department of Human Communication Sciences, University of Sheffield, 362 Mushroom Lane, Sheffield, S10 2TS, United Kingdom

art ic l e i nf o

a b s t r a c t

Article history: Received 10 September 2012 Received in revised form 4 October 2013 Accepted 24 October 2013 Available online 31 October 2013

One factor in syntactic impairment in aphasia might be damage to general structure processing systems. In such a case, deficits would be evident in the processing of syntactically structured non-linguistic information. To explore this hypothesis, we examined performances on artificial grammar learning (AGL) tasks in which the grammar was expressed in non-linguistic visual forms. In the first experiment, AGL behavior of four aphasic participants with severe syntactic impairment, five aphasic participants without syntactic impairment, and healthy controls was examined. Participants were trained on sequences of nonsense stimuli with the structure AnBn. Data were analyzed at an individual level to identify different behavioral profiles and account for heterogeneity in aphasic as well as healthy groups. Healthy controls and patients without syntactic impairment were more likely to learn configurational (item order) than quantitative (counting) regularities. Quantitative regularities were only detected by individuals who also detected the configurational properties of the stimulus sequences. By contrast, two individuals with syntactic impairment learned quantitative regularities, but showed no sensitivity towards configurational structure. They also failed to detect configurational structure in a second experiment in which sequences were structured by the grammar A þ B þ . We discuss the potential relationship between AGL and processing of word order as well as the potential of AGL in clinical practice. & 2013 Elsevier Ltd. All rights reserved.

Keywords: Aphasia Agrammatism Syntax Artificial grammar learning

1. Introduction The ability to parse sequences, detect regularities and generalize them onto new sequences is fundamental to language processing. A number of experiments have revealed a capacity to extract syntactic information after brief exposure to structured stimulus sequences in infants (Marcus, Vijayan, Bandi Rao, & Vishton, 1999; Saffran et al., 2008), and adults (Pothos, 2007). Sensitivity to sequential regularities has been demonstrated across sensory modalities and when stimuli contained no lexical-semantic information. In the current study we explored the relationship between syntactic performance in aphasia and sensitivity to syntactic structure in sequences of meaningless visual stimuli. People with aphasic syntactic impairment have difficulties in production of well-formed sentence structures, but also impaired understanding of sentences. When presented with semantically reversible sentences like The cat that the dog is biting is black or The dog that the cat is biting is black, they may be unable to determine “who did what to whom” (Berndt, Mitchum, & Haendiges, 1996; Caplan, Baker, & Dehaut, 1985; Caramazza & Zurif, 1976). In the n

Corresponding author. Tel.: þ 44 20 7679 24270. E-mail addresses: [email protected] (V.C. Zimmerer), [email protected] (P.E. Cowell), [email protected] (R.A. Varley). 0028-3932/$ - see front matter & 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.neuropsychologia.2013.10.014

case of mild or moderate syntactic aphasia, an individual might use word order or semantic–heuristic strategies to decode structural/functional information encoded within the sentence structure. This may result in correct interpretation of structures like canonical transitive actives (The lion kills the man) using a configurational ‘agent-first’ heuristic. In the case of severe syntactic impairment, the capacity to use configurational information might also be disrupted and performance in interpreting reversible sentences will be unsuccessful in canonical actives as well. Sentence comprehension and production require multicomponent cognitive processing, and agrammatic behavior might arise from failure of one or more of these components and their interactions. Since agrammatism was first described, there has been a range of attempts to account for underlying cognitive failures (Caramazza, Capitani, Rey, & Berndt, 2001; Martin, 2006). Proposals commonly focus on mechanisms specifically related to the interpretation of natural language sentences, such as word order transformation (Drai & Grodzinsky, 2006; Grodzinsky, 2000), lexical processing (Biassou, Obler, Nespoulous, Dordain, & Harris, 1997; Druks, 2002; Friederici, 1982), or thematic role assignment (Saffran, Schwartz, & Linebarger, 1998; Wassenaar & Hagoort, 2007). Others suggest that impaired linguistic working memory systems limit the ability to retain or manipulate linguistic information (Caplan & Waters, 1999; Caplan, Waters, DeDe, Michaud, & Reddy, 2007; Haarmann, Just, & Carpenter, 1997; Haarmann & Kolk, 1994;

26

V.C. Zimmerer et al. / Neuropsychologia 53 (2014) 25–38

Miyake, Carpenter, & Just, 1994). However, if pre-verbal infants and healthy adults are able to detect syntactic regularities in sequences of nonsense stimuli, this suggests that general structure processing mechanisms may be a core element of syntactic behavior. With regard to agrammatism, investigations of natural language sentence processing inevitably confound structural processing capacity with issues of lexical-semantic ability. A key question is whether agrammatic behavior can be identified in the processing of sequences of non-lexical and non-semantic stimuli. The artificial grammar learning (AGL; Reber, 1967) paradigm allows investigations of syntactic cognition in absence of linguistic and, in particular, lexical-semantic information. The typical AGL experiment consists of two phases: in the training phase, participants are presented with sequences of nonsense stimuli which are generated by a set of grammatical rules (the target grammar). In the following test phase, new stimulus sequences are presented which are either consistent or inconsistent with the target grammar. Participants decide whether test sequences “fit” the training sequences. Without receiving any explicit training regarding the syntactic structure of grammatical stimuli, participants spontaneously establish their own acceptance/rejection criteria for the test sequences. Their decision patterns allow insight into the syntactic properties of grammatical sequences that were detected. Studies on healthy speakers have reported correlations between AGL performance and language processing abilities (Conway & Pisoni, 2008; Misyak & Christiansen, 2012). Misyak, Christiansen, and Tomblin (2010) reached similar conclusions using an AGL paradigm which incorporated reaction time measurements. Functional brain imaging studies have shown that AGL tasks evoke increased activation in left inferior frontal lobe areas thought responsible for grammatical processing, regardless of whether stimuli are letters (Petersson, Forkstam, & Ingvar, 2004; Petersson, Folia, & Hagoort, 2012), written syllables (Bahlmann, Schubotz, & Friederici, 2008) or nonsense shapes (Bahlmann, Schubotz, Mueller, Koester, & Friederici, 2009). These areas are typically lesioned in people with agrammatic aphasia. Previous patient reports describe “agrammatic AGL” whereby groups with syntactic impairment performed worse than healthy controls in discriminating grammatical from ungrammatical sequences. These studies investigated configurational processing, i.e., processing of order in a sequence. Dominey, Hoen, Blanc and Lelekov-Boissard (2003) trained seven agrammatic participants on the structure 123213, in which order transformation of the first half of the sequence (123) resulted in the second half (213). The structure was mapped onto letters, resulting in sequences like ABCBAC or DEFEDF. Participants were trained on ten grammatical letter sequences and then asked to classify twenty new sequences as correct or incorrect. The aphasic group performed poorly. However, performance improved when the simpler target grammar 123123 was used, and the second half of the sequence was a repetition of the first (e.g., ABCABC). AGL performance correlated with sentence comprehension scores, with association between performance on the difficult grammar and comprehension of non-canonical sentences, and performance on the easier grammar and comprehension of canonical sentences. Dominey et al. argue that sequential order transformation is necessary for sentences which are derived from canonical structures. For instance, the comprehension of relatives (It was the ball that the cat chased) would require order transformation from the canonical active form (The cat chased the ball). Their results suggest that agrammatism represents an impairment of a transformational principle that is expressed across different informational domains. A fundamental part of configurational processing is the detection of order, which determines semantic roles even in simple structures such as the English canonical active. Most AGL studies use finite-state grammars which involve no order transformation

and their results show that healthy participants can detect how sequences start, end, and which stimuli can appear next to each other. Christiansen, Louise Kelly, Shillcock, and Greenfield (2010) trained seven aphasic participants on sequences generated by a finite-state target grammar. Stimuli were simple geometric shapes. In the training phase participants were presented with series of paired sequences and had to judge whether they were identical. They were then told that the sequences had been generated by a set of rules. In the test phase, participants encountered 39 new sequences (19 of which were violations of the target grammar) and judged whether they were grammatical or not. The aphasic group was significantly less successful in distinguishing grammatical from ungrammatical test sequences than healthy controls matched for non-verbal intelligence. Christiansen et al. not only investigated the effect of grammaticality, but also the surface familiarity of the novel sequences to training sequences. While grammaticality significantly affected the behavior of the control group, a marginally significant trend indicated that aphasic behavior could have been driven by the anchor strength of the test sequences, i.e., the frequency with which the first and final stimuli in a sequence appeared at the same position during training. In the current study we further examine the relationship between agrammatism and AGL performance and address two issues which have not been investigated previously. While the two previous studies of aphasic AGL performance (Christiansen et al., 2010; Dominey et al., 2003) compared findings from an aphasic group to those of non-impaired controls, it is not clear whether the behaviors identified in the patients were due specifically to syntactic impairment, or rather aphasia or brain damage more generally. Evaluation of AGL behavior also needs to take into account the heterogeneity of participant groups. Analysis of AGL data from healthy individuals has revealed substantial variations in behavior (Visser, Raijmakers, & Pothos, 2009; Zimmerer, Cowell, & Varley, 2011). Some healthy participants fail to detect any task-relevant properties of stimulus sequences, resulting in chance performance, while others perform at ceiling. Furthermore, participants with a similar number of correct responses systematically rejected different types of violation, suggesting that they used different grammaticality criteria in making decisions. Withingroup variability is likely to be more marked in investigations of aphasia. Profiles of syntactic behavior differ even among patients with lesions at similar locations (Berndt & Caramazza, 1999; Caramazza et al., 2001). AGL experiments, especially with clinical populations, should therefore not only consider the total number of “correct” decisions consistent with the target grammar, but also which types of test sequences each individual participant consistently accepts or rejects. In the current study we compare performance profiles of aphasic individuals against the range of individual profiles in control participants. We examined the individual performance profiles of two groups of aphasic participants (with and without syntactic impairment) and healthy controls on an AGL task. We adopted a case series design in which the performance of individuals with syntactic impairment was compared with the range of individual performances found in controls. Participants with syntactic impairment were identified through chance performance on both spoken and written reversible sentence comprehension tests, in actives as well as passives, and an absence of productive syntactic capacity in spoken and written output. Within the syntactically impaired group some participants showed residual lexical processing, while one showed marked lexical comprehension impairment and might typically be described as globally aphasic. Experiment 1 used the target grammar AnBn. In sequences generated by this non-finite grammar, a number of symbols of class A are followed by the same number of stimuli of class B


(e.g., AABB, AAABBB, etc.). The grammar has been used in a number of AGL studies, and interpretation of how participants represent the structure has changed over time. In early psycholinguistic literature (e.g., Chomsky, 1957) it is described as a centerembedding phrase-structure grammar. A “sentence” can nest another sentence of the same structure in its center, S-A[S]B (e. g., [A[A[AB]B]B]). Earlier AGL studies utilizing the grammar have adopted this interpretation (Fitch & Hauser, 2004; Friederici, Bahlmann, Heim, Schubotz, & Anwander, 2006; Hauser, Chomsky, & Fitch, 2002). However, this view has been criticized on the basis that participants might simply count the number of stimuli of class A and B (de Vries, Monaghan, Knecht, & Zwitserlood, 2008; Hochmann, Azadpour, & Mehler, 2008; Perruchet & Rey, 2005; Pinker & Jackendoff, 2005; Zimmerer et al., 2011), and this criticism has been accepted by former proponents of the center-embedding account (Bahlmann et al., 2008; Fitch & Friederici, 2012; Hauser, Barner, & O'Donnel, 2007). We adopt this latter view and assume that grammatical AnBn sequences are defined by configuration (all items of one class precede the other class; items of class A appears first), as well as quantities (the numbers of As and Bs match). Processing of configurational information is essential to syntactic processing. Counting, however, is not considered relevant to syntax. The inclusion of these different types of patterns brings with it the advantage of having a “control task” within the experiment. The non-linguistic quantitative pattern makes it possible to detect learning even if participants are insensitive to configuration, and thus verify that instructions and task were understood. The performance of healthy participants in learning AnBn sequences is well established. Participants are better at detecting configuration than quantity (de Vries et al., 2008; Hochmann et al., 2008). Zimmerer et al. (2011) investigated the range of individual AGL behavior in young healthy participants exposed to AnBn sequences. Two different groups were tested on auditory or visual stimuli. The participant groups' mean percentage of correct judgments was comparable to means reported in previous studies (e.g., Fitch & Hauser, 2004), however, with a considerable range (50–98%). While range has not been reported in previous studies using this grammar, Misyak and Christiansen (2012) also detected considerable within-group range. Further investigation revealed a number of individual performance profiles (Zimmerer et al., 2011). Most healthy participants detected structural properties of sequences, however, for a small number there was no evidence of syntactic learning. Based on the types of violations rejected it is possible to characterize the decision criteria applied by individuals. Every participant who showed evidence of learning at least learned that: (1) All items of one class appear together in one block, uninterrupted by another class. (1) describes individuals who consistently rejected sequences like ABBABA or BABAAB, but accepted sequences structured BnAn (e.g., BBBAAA) as well as sequences with a violation of item number (e.g., AABBB). In addition to (1), participants could detect none, one or both of the following: (2) Sequences start (or are anchored) with one stimulus of class A and end with one stimulus of class B. (3) The number of items of the classes A and B match. In summary, participants who showed evidence for learning detected only configurational properties, or configurational as well as quantitative properties. None of the participants detected the quantitative pattern without identifying configurational relations, which suggests that recognition of configurational relations is more fundamental to pattern detection than counting. In the current study, it was considered likely that patients would show

27

an overall poorer performance than healthy controls, regardless of syntactic impairment, as the result of brain damage. We did not make hypotheses based on the overall number of correct judgments. Rather, we investigated whether the behavioral profiles of patients with syntactic impairment, but not the ones of patients without syntactic impairment, would deviate from the profiles observed in healthy participants. Experiment 2 used the finite-state grammar A þ B þ , which generates sequences in which a number of stimuli of class A are followed by a number of stimuli of class B (e.g., AAABB, AAABBB, AABBBB). The numbers do not have to match for a sequence to be grammatical. In learning A þ B þ , only the detection of configurational relations is necessary to distinguish grammatical from ungrammatical sequences. The grammar was chosen to remove non-syntactic quantitative patterns from the grammar and narrow the learner's attention to configuration. We report data from participants with syntactic impairment. Zimmerer et al. (2011) showed that in healthy participants, AGL performance was similar regardless of whether stimuli were auditory (spoken non-word syllables) or visual (nonsense shapes). Because the participants with syntactic impairment in this study had extensive lesions involving the left superior temporal gyrus, resulting in impairment of auditory-phonological processing, all experiments in this report used visual stimuli. Experiments on healthy controls had approval from the University of Sheffield Research Ethics Committee. Ethical approval for experiments on participants with neurological damage was granted by the local NHS Research Ethics Committee (08/H1308/ 32). Experiments were written using DMDX (Forster & Forster, 2003) and run on a Dell Inspiron 5150 laptop PC.

2. Experiment 1 2.1. Participants 2.1.1. Participants with syntactic impairment Four severely aphasic men SA, SO, PR and JB took part in both AGL experiments. All were native speakers of English, and were aged between 61 and 71 years. SO, PR and JB had primary vascular lesions in the left middle cerebral artery territory. SA had a subdural empyema in the left sylvian fissure which resulted in secondary vascular lesion. All participants had damage to left perisylvian language zones, including BA 44, and were severely aphasic (Fig. 1 for structural scans of SA, SO and PR). All four showed little evidence of productive use of phrase or clause structures in spoken and written language, although PR was able to produce a small range of fully lexicalized constructions (e.g., I don't know) and sentence connectives (e.g., …and then…; …but…) which he used between gestures. Comprehension of spoken language was impaired in all cases. Reading comprehension was also impaired, although less severely than the auditoryphonological channel. JB also displayed profound deficits in lexical processing and was classified as globally aphasic, while SA, SO and PR were classed as having severe agrammatic aphasia. All participants were pre-morbidly right-handed, with the exception of PR who was left-handed. Since PR's aphasia was caused by a lesion to the left-hemisphere, he belonged to the majority of left-handers who are left-hemisphere dominant for language (Pujol, Deus, Losilla, & Capdevila, 1999). All participants except SA suffered from right-sided hemiparesis. SA had a sensory deficit on his right side that limited use of his right hand, so that he preferred using his left hand. All participants were at least in their third year post-onset at the time of testing. Table 1 provides the scores of the syntactically impaired participants on a range of spoken and written language tests.

28


Fig. 1. Structural brain scans for SA, SO and PR (from left to right panel).

Table 1 Results of linguistic and non-linguistic assessment of the four participants with syntactic impairment. Asterisks mark above chance performance (po .05). Participant Age (years post-onset)

SA 62 (15)

SO 61 (8)

PR 63 (8)

JB 71 (3)

Lexical-semantic assessments ADA spoken word picture matching (chance ¼ 16.5) ADA written word picture matching (chance ¼ 16.5) ADA spoken synonym matching (chance ¼ 80) ADA written synonym matching (chance ¼ 80) PALPA 54 spoken picture naming PALPA 54 written picture naming

60/66n 62/66n 123/160n 121/160n 0/60 24/60

61/66n 66/66n 121/160n 145/160n 0/60 0/60

61/66n 66/66n 121/160n 145/160n 0/60 2/60

54/66n 52/66n 78/160 114/160n 0/60 0/60

Syntactic assessments Comprehension of spoken reversible sentences (chance ¼50) Comprehension of written reversible sentences (chance ¼ 50) Written grammaticality judgments (chance ¼20)

49/100 42/100 26/40n

47/100 43/100 35/40n

38/100 49/100 21/40

47/100 50/100 24/40

Verbal working memory assessment PALPA 13 digit span (recognition)

3 items

5 items

4 items

1 item

Non-linguistic assessments Visual Patterns Test (percentile for age group) Ishihara color test

11.5 (90th) 14/14

– –

8.6 (40th) 13/14

– 14/14

Lexical-semantic assessments were taken from the Action for Dysphasic Adults (ADA) Auditory Comprehension Battery (Franklin, Turner, & Ellis, 1992) and the Psycholinguistic Assessments of Language Processing in Aphasia (PALPA; Kay, Lesser, & Coltheart, 1992). Word-picture matching tests required a stimulus word to be matched to a corresponding picture in the presence of visual, phonological/orthographic and semantic distracters. While word-picture matching tests assess the comprehension of concrete words, the synonym matching tests also included words of low imageability. Picture naming tests measured both spoken and written lexical production. Syntactic assessments were designed for the purposes of the experiment. Reversible sentence tests required the participant to match a spoken or written sentence to a picture representing the described event. One image depicted the event, while a second distracter image showed the event with reversed roles of the protagonists. Active and passive forms were tested. In one trial for example, the participant would be presented with one of the four sentences The man kills the lion, The lion kills the man, The lion is killed by the man or The man is killed by the lion together with two pictures, one of a lion killing a man and one of a man killing a lion. A written grammaticality judgment test required grammatical/ ungrammatical decisions. Distractor items included severe violations such as a misplaced or missing VP. For all participants performance on reversible sentences was poor on both actives and passives in spoken and written

modalities. All were impaired on grammaticality judgments, with three of the patients showing marked impairment, while SO demonstrated some residual ability on this test. Lexical production was impaired in all cases, with the exception of SA who could produce written names for 24 of the 60 picture stimuli. On the basis of the profile of test scores, participants SA, SO, and PR were classified as agrammatic, while JB was classified as globally aphasic due to impaired lexical-semantic comprehension in addition to poor syntactic performance. Phonological working memory was tested with PALPA test 13. Digit span tests commonly involve the spoken recall of digits. However, as participants had severe difficulties with language production, a recognition paradigm was used. Participants heard two digit sequences and had to indicate whether they matched (e.g., 3–5–7; 3–5–7). The average span for recognition in an agecontrol group was 8 items (Ankerstein, Peace, Evans, & Varley, unpublished). Aphasic performance ranged from 1 to 5 items, and was thus below average. Visual working memory was tested with the Visual Patterns Test (VPT; Della Sala, Gray, Baddeley, & Wilson, 1997), which requires participants to recall grid patterns of increasing complexity. VPT scores were available for two patients. SA scored at the 90th percentile in relation to adults from his age range with no neurological impairment. PR scored at the 40th percentile. Since visual AGL stimuli were classified by color, participants were tested on the Ishihara Test Chart for Color Deficiency


29

Table 2 Aphasic participants without syntactic impairment and their sentence comprehension scores. Participants were coded by age and sex (42M ¼ male, age 42). Participant

42M

51M

65M1

65M2

84F

Years post onset Comprehension of spoken reversible sentences (chance ¼10)

3 18/20

1 20/20

2 20/20

5 20/20

3 19/20

(Ishihara, 2007). No participant was diagnosed as color blind. SO was not available for testing. None of the participants had a marked hemianopia. 2.1.1.1. Aphasic participants without syntactic impairment. Five nonagrammatic controls with left-hemisphere lesions and chronic aphasia were recruited. These patients were selected from an established database of aphasic participants on the basis of having ceiling or near-ceiling scores on a screening version of the spoken reversible sentences test. This test was similar in format to the more extended version used to probe syntactic comprehension in the agrammatic patients (e.g., selection of the target from two pictures displaying reversed roles of the protagonists, inclusion of active and passive sentences) (Table 2). All were native speakers of English. Four were male, one was female. All were aged between 42 and 84 years (mean age 61.4). All were at least in the second year post-onset. 2.1.1.2. Healthy controls. Ten healthy participants aged 50–78 years (mean age 62; 2 male, 8 female) were recruited as normative controls. All participants were native speakers of English, and none reported history of neurological impairment. 2.1.2. Stimuli and procedure Participants took part in an AGL experiment that employed the grammar AnBn expressed in visual stimuli. The training phase contained 72 sequences generated by the grammar. The value of n was 2, 3, or 4, resulting in sequences with the length of 4, 6, or 8 stimuli. There were 24 sequences for each different n. Sequences were presented in three trial blocks, with a self-timed break between blocks. The test phase contained a total of 120 novel sequences. Half had the structure AnBn (grammatical). The others violated the grammar (ungrammatical). There were five types of violation sequences: 1. Type 1 (quantitative violation; e.g., AABBB) had an uneven total number of items in the sequence. The number of As and Bs did not match. 2. Type 2 (quantitative violation; e.g., AAAABB) did not have matching numbers of As and Bs. 3. Type 3 (configurational violation; e.g., BBBAAA) had an incorrect initial and final item. It contained the illegal bigram BA. 4. Type 4 (configurational violation; e.g., BABAAB) had an illegal initial item and contained the illegal bigram BA. Items of one class did not appear together. 5. Type 5 (configurational violation; e.g., ABBABA) had an illegal final item and contained the illegal bigram BA. Items of one class did not appear together. There were 12 sequences for each violation type. Grammatical and ungrammatical test sequences were coded according to familiarity measures (Knowlton & Squire, 1994; Redington & Chater, 1996). Associative Chunk Strength (ACS) describes the average frequency with which bigrams and trigrams appeared in the training set. Anchor Strength (AS) denotes the average frequency with which the initial and final stimulus bigrams and trigrams appeared at the respective positions during

training. Chunk Novelty (CN) counts the number of bigrams and trigrams in a sequence which did not appear during training. Independent samples t-tests showed that grammatical and ungrammatical test sequences did not differ significantly in ACS, t(118) ¼ 1.153, p ¼.251, or CN, t(118) ¼1.255, p ¼.212. Grammatical sequences had significantly higher AS than ungrammatical sequences, t(118) ¼ 6.096, p o.001. However, note that sequences of violation types 3–5 had novel anchors by design. As in the training phase, test phase sequences were presented in three trial blocks, with a self-timed break between each block. The order of the sequences in both phases was randomized using Microsoft Excel. In the test phase, no more than four consecutive trials were either all grammatical or ungrammatical. Sequence presentation order was the same for each participant. For each A and B, one of eight nonsense shapes was presented. Stimuli for class A were blue and rounded. Stimuli for class B were red and angled (Fig. 2). Each shape was approximately 6 6 cm. Sequences were presented one at a time in the middle of the screen on a light gray background. Each stimulus was presented for 400 ms, with an interval of 200 ms between stimuli. In the learning phase, there was an interval of 4.5 s between each sequence during which a fixation cross was presented. In the test phase, participants made decisions by pressing one of two buttons on a computer mouse. The button for accepting sequences was marked green, the one for rejecting sequences was marked red. If no response was given within 4.5 s, the program moved on to the next sequence. For healthy participants, and aphasic, non-agrammatic controls, written instructions were presented on the screen. Before the training phase, they were: “Just relax and watch.” Before the test phase, they were the following: “You will now see new shapes. Decide whether what you saw fits to the examples you saw before. Decision time is important, so decide as quickly as possible.” Due to severe comprehension impairment, instructions differed for participants with syntactic difficulties. Written instructions were supported by spoken and gestural cues. Before the training phase, participants with syntactic impairment were asked to watch the screen. Before the test phase, they were given simplified instructions on paper in a 26pt font: “Now you will see some more shapes. Do they follow the pattern of the shapes you just saw? GREEN button – YES – same pattern. RED button – NO – different pattern. Decide quickly.” Each sentence was presented in a different line on the paper. The words “green” and “red” were in green and red font, respectively. In addition, the experimenter provided oral instructions and gestures. Participants were told that there would be new shapes and that some would be good (thumb up) and some bad (thumb down). The experimenter also pointed at the mouse buttons for “good” and “bad”. The test phase started when the participant indicated that he understood what was requested of him.

2.1.3. Data analysis Data for the healthy control sample (n ¼10) were analyzed using multiple regression in a manner similar to Christiansen et al. (2010). The aim was to determine the effects of familiarity measures and grammaticality (i.e., adherence to the target grammar) on rejection of sequences in the test phase. Each of

30


Fig. 2. Stimuli used in the study. Stimuli of class A were blue and had a rounded shape (top two panels). Stimuli of class B were red and had an angled shape (bottom two panels). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

the 120 test sequences was coded according to the predictors ACS, AS, CN and grammaticality. The outcome variable was the percentage of participants who rejected a given sequence. This analysis was not conducted for either patient group due to the small sample sizes. A primary objective of our study was to investigate individual performance patterns from controls and patients. In the first step of this analysis, we compared the effects of ACS, AS, CN and the categorical variable Sequence Type (with six categories: Violation types 1–5 and grammatical sequences). By contrast to the group data from healthy participants, the outcome variable for individuals was binary: a single participant either accepted or rejected a sequence. As a result of this binary outcome and the combination of different predictor types, data were not suited to linear regression analysis. We therefore carried out logistic regressions based on each individual's test phase responses. For the variable Sequence Type we used grammatical sequences as a reference category. One issue was that some participants either accepted or rejected all sequences of one type. This behavior resulted in empty or full categories which made it impossible for the logistic regression analysis to fully converge. In these cases, our models assumed one added or subtracted rejection (e.g., if a participant accepted all 60 grammatical sequences, the model added one

rejection). The application of logistic regression to within-subject profile analysis is not a conventional application of the technique. However, its adaptation in this instance allowed the data to be statistically framed in a manner that closely mirrored the experimental design and was faithful to the study's objectives to examine individual performance patterns. In the second step of investigation of individual data, we used a template analysis (Zimmerer et al., 2011) to determine different behavioral profiles on the basis of effects of sequence type. In this type of analysis, responses are compared to those generated by a number of hypothesized decision patterns, each of which suggested the learning of different syntactic features. Each decision pattern is defined by types of sequences accepted and rejected. For instance, a participant sensitive to configurational information in AnBn sequences, but not quantitative patterns, would tend to reject violations of types 3–5, but accept violations of type 1 and 2, as well as grammatical sequences. Conversely, a participant sensitive to quantitative patterns, but not to configuration, would reject violations of type 1 and 2, and accept violations of types 3–5, as well as grammatical sequences. Each decision pattern therefore yields a different template for analysis. Table 3 describes the eight templates that were considered in this experiment. For each template, a D score (Perruchet & Pacteau, 1990; Zimmerer et al., 2011) was


31

Table 3 Templates for the analysis of Experiment 1 (numbers represent violation types)a. Template

Criteria for grammaticality

Ungrammatical sequences according to template

Grammatical sequences according to template

AnBn EVEN EQUAL A(A þB)n (A þ B)nB AþBþ (A þ B þ ) þ(B þ A þ ) (AnBn) þ (BnAn)

A number of As is followed by an equal number of Bs Total number of items in a sequence is even Equal number of As and Bs Sequence starts with A Sequence ends with B All As precede the Bs All items of one class precede all items of the other class All items of one class precede all items of the other class, equal number of As and Bs

1–5 1 1, 2 3, 4 3, 5 3–5 4, 5 1, 2, 4, 5

AnBn 2–5, AnBn 3–5, AnBn 1, 2, 5, AnBn 1, 2, 4, AnBn 1, 2, AnBn 1–3, AnBn 3, AnBn

a Notation: “An”: n instances of A “An”: zero or more instances of A “A þ ”: one or more instances of A “A þB”: either A or B.

calculated using the following formula: D score ¼ ðpercentage of ungrammatical sequences rejectedÞ

Table 4 Regression coefficients for healthy controls. Asterisks denote p o .001.

–ðpercentage of grammatical sequences rejectedÞ As opposed to d′, D scores are normalized and allow comparison between templates with different numbers of sequences to be accepted and rejected. D scores range from 100 to 100, with zero indicating that grammatical and ungrammatical sequences were rejected at equal percentages. For participants whose data showed a significant effect of sequence type according to the logistic regression, the template that was above chance and yielded the highest D score was considered the “best template” and therefore represented the decision criteria most likely applied in the test phase. 2.1.4. Results 2.1.4.1. Group comparisons. Mean proportion of correct decisions (acceptance of AnBn test sequences, rejection of violation sequences) was 64.75% (SEM¼2.9) for aphasic participants with syntactic impairment, 70.8% (SEM¼5.39) for aphasic participants without syntactic impairment, and 88.5% (SEM¼4.96) for healthy participants. A One-way ANOVA revealed a significant effect of group, F(2,16)¼ 5.66, p¼.014. Comparisons between groups showed that the main effect was due mainly to the healthy control group making more correct decisions than the group with syntactic disorder, p¼.009. The difference was also significant according to a Bonferroni adjusted threshold (.017). The difference between healthy participants and aphasic participants without syntactic impairment was significant, p¼ .029, though not at corrected levels. As demonstrated in previous studies (Visser et al., 2009; Zimmerer et al., 2011), similarities at such a level do not imply that participants applied the same decision criteria. We therefore carried out analyses at individual level. 2.1.4.2. Healthy controls. Older healthy controls rejected 8.87% (SEM¼ 4.32) of the grammatical sequences. Rejection of ungrammatical sequences differed between violation types. The group rejected 78.33% (SEM¼7.68) of the sequences of violation type 1, 87.8% (SEM¼5.13) of violation type 2, 85% (SEM¼6.06) of violation type 3, 93.33% (SEM¼4.27) of violation type 4, and 86.52% (SEM¼6.84) of violation type 5. Paired sample t-tests showed that violations of type 1 were rejected less often than violations of type 2 (p ¼.043) and 4 (p¼.012). Violations of type 3 were rejected less often than violations of type 4 (p¼ .023). However, these differences did not withstand Bonferroni corrections (adjusted threshold: po.005). We carried out a multiple regression with ACS, AS, CN and grammaticality as predictors and rate of

Constant Associative Chunk Strength Anchor Strength Chunk Novelty Grammaticality

B

SE B

β

.244 .024 .007 .007 .727

.058 .033 .039 .007 .029

.028 .007 .036 .939nn

rejections as the outcome (see Section 2.1.3). There was a significant effect of grammaticality (Table 4), but not ACS, AS or CN. Logistic regressions carried out for each participant showed significant effects of sequence type for each participant except 73M (Appendix D). Individual acceptance/rejection patterns of healthy controls are presented in Table 5. Individual data were analyzed using the template analysis. Appendix A contains all template scores for each individual. For nine participants, the template with the highest D score was AnBn. Their mean D score for AnBn was 86 (range: 66.38–100). 2.1.4.3. Aphasic participants without syntactic impairment. Aphasic non-agrammatic controls rejected 22.94% (SEM ¼8.33) of the grammatical sequences. Rejection of ungrammatical sequences differed between violation types. The group rejected 37.45% (SEM¼12.85) of the sequences of violation type 1, 41.97% (SEM¼16.13) of violation type 2, 83.33% (SEM ¼6.97) of violation type 3, 83.18% (SEM ¼5.84) of violation type 4, and 77.06% (SEM¼8.33) of violation type 5. Individual data are presented in Table 6. Logistic regressions carried out for each individual participant revealed significant effects of Sequence Type for all participants except 65M2 (see Appendix D). Appendix B contains template scores of all participants. The best template for 65M1 was AnBn with a D score of 76.67, as he consistently rejected configurational and quantitative violations. The three participants 42M, 51M and 84F consistently rejected configurational violation sequences in which a stimulus of class A followed a stimulus of class B. Their best template was A þ B þ . 2.1.4.4. Participants with syntactic impairment. Syntactically impaired participants rejected 23.12% (SEM ¼7.47) of the grammatical sequences. They rejected 71.59% (SEM ¼13.14) of the sequences of violation type 1, 75% (SEM ¼17.01) of violation type 2, 25.95% (SEM ¼9.61) of violation type 3, 40.72% (SEM ¼ 23.12) of violation type 4, and 45.83% (SEM ¼17.81) of violation

32


Table 5 Rejection patterns across all test phase sequence types for healthy participants. Numbers present percentage of rejected sequences from a given type. AnBn: sequences in the target grammar (e.g., AAABBB). Viol. type 1: illegal total number of items (e.g., AABBB). Viol. type 2: number of As and number of Bs mismatch (e.g., AAAABB). Viol. type 3: BnAn (e.g., BBBAAA). Viol. type 4: illegal initial item, items of one class do not appear together (e.g., BABAAB). Viol. type 5: illegal final item, items of one class do not appear together (e.g., ABBABA). (Viol.¼ violation type). Sequence type

50F 53F1 53F2 58F 60F 62F 66M 67F 73M 78F

AnBn (grammatical)

Viol. 1 (quantity)

Viol. 2 (quantity)

Viol. 3 (configuration)



0 0 0 3.39 5.17 16.77 5 0.34 44.83 3.33

100 75 100 83.33 66.67 75 100 68.33 25 100

100 66.67 100 100 81.82 91.67 100 83.33 54.55 100

100 100 100 100 75 75 75 83.33 41.67 100

100 100 100 100 83.33 91.67 100 100 58.33 100

100 91.67 100 100 66.67 81.82 100 100 33.33 91.67

Table 6 Rejection patterns across all test phase sequence types for aphasic participants without syntactic impairment. Numbers present percentage of rejected sequences from a given type. AnBn: sequences in the target grammar (e.g., AAABBB). Viol. type 1: illegal total number of items (e.g., AABBB). Viol. type 2: number of As and number of Bs mismatch (e.g., AAAABB). Viol. type 3: BnAn (e.g., BBBAAA). Viol. type 4: illegal initial item, items of one class do not appear together (e.g., BABAAB). Viol. type 5: illegal final item, items of one class do not appear together (e.g., ABBABA). (Viol.¼ violation). Sequence type

42M 51M 65M1 65M2 84F

AnBn (grammatical)

Viol. 1 (quantity)

Viol. 2 (quantity)




18.52 13.34 15 55.93 18.52

27.27 16.67 75 60 8.33

18.18 33.33 100 50 8.33

75 100 100 75 66.67

66.67 75 100 83.33 90.91

80 91.67 83.33 60 75

Table 7 Rejection patterns across all test phase sequence types for participants with syntactic impairment. Numbers present percentage of rejected sequences from a given type. AnBn: sequences in the target grammar (e.g., AAABBB). Viol. type 1: illegal total number of items (e.g., AABBB). Viol. type 2: number of As and number of Bs mismatch (e.g., AAAABB). Viol. type 3: BnAn (e.g., BBBAAA). Viol. type 4: illegal initial item, violation of block consistency (e.g., BABAAB). Viol. type 5: illegal final item, violation of block consistency (e.g., ABBABA). (Viol.¼ violation). Sequence type

SA SO PR JB

AnBn (grammatical)

Viol. 1 (quantity)

Viol. 2 (quantity)




8.33 16.67 43.33 24.14

75 100 75 36.36

91.67 100 83.33 25

0 25 33.33 45.45

0 8.33 54.55 100

8.33 41.67 41.67 91.67

Table 8 D scores of all eight templates for participants with syntactic impairment. Numbers in bold signify the “best template”, i.e. the highest D score of an individual participant whose behavior was significantly affected by sequence type. JB's behavior was affected by sequence type at a marginally significant level. AnBn: A number of As is followed by an equal number of Bs. EVEN: Total number of items in a sequence is even. EQUAL: Equal number of As and Bs. A(A þB)n: Sequence starts with A. (A þB)nB: Sequence ends with B. A þ B þ : All As precede the Bs. (A þ B þ ) þ(B þ A þ ): Items of one class appear together. (AnBn) þ (BnAn): Items of one class appear together, equal number of As and Bs. Template D score

SA SO PR JB

AnBn

EVEN

EQUAL

A(A þ B)n

(A þ B)nB

AþBþ

(A þ B þ ) þ (B þ A þ )

(AnBn)þ (BnAn)

26.67 38.33 14.29 35.51

59.26 71.30 27.34 5.94

77.08 80.21 36.01 14.13

27.08 23.96 8.61 38.32

21.88 3.13 16.18 34.78

26.98 15.48 10.71 53.49

21.88 13.54 3.22 67.39

36.81 44.44 22.16 35.51

type 5. Individual data are reported in Table 7. Logistic regressions carried out for each participant revealed significant effects of sequence type for SA and SO, with a marginal trend for JB (p ¼ .067) (see Appendix D).

Individual template scores are reported in Table 8. SA consistently rejected violations of quantitative patterns, but accepted configurational violations. The template analysis determined EQUAL as the best template, with a D score of 77.08. SO showed


33

no consistent pattern of response criterion according to the analyses employed in this experiment. This behavior, too, was observed in healthy participants. 2.2. Experiment 2

Fig. 3. Individual AGL performance as depicted as a function of best fit in the grammatical template analysis (Table 2). Four participant groups are displayed: Younger healthy controls as reported in Zimmerer et al. (2011), older healthy controls, aphasic, non-agrammatic participants and participants with syntactic impairment. Syntactically-impaired participants are labeled. Chance level performances were excluded.

similar behavior to SA. His best template was also EQUAL, with a D score of 80.21. PR consistently rejected violations of quantitative patterns. However, his behavior on other violations was not consistent. JB consistently rejected sequences in which stimulus classes interchanged several times. His highest D score was 67.39 on the template (A þ B þ )þ (B þ A þ ).

2.1.5. Discussion Comparison at group level showed that healthy participants rejected more violations of AnBn than aphasic groups. However, analysis of individual decision criteria (summarized in Fig. 3) suggests that the behavior of healthy participants and aphasic participants without syntactic impairment falls within the normative pattern of individual AGL performance described by Zimmerer et al. (2011; see above). Nine out of ten healthy participants detected both configurational and quantitative patterns; one healthy participant detected no structure. Four aphasic participants without syntactic impairment either detected configurational patterns, or both configurational and quantitative patterns; one detected no structure. By contrast, two agrammatic participants, SA and SO, showed behavior that substantially deviated from this normal pattern. Both consistently rejected violations of quantitative patterns (suggesting that they understood task instructions), but did not use configurational information. They detected that the number of A and B stimuli matched in grammatical sequences, but were insensitive to the position in which stimuli appeared, regardless of familiarity as captured by ACS, AS and CN. Given that people without syntactic impairment are more likely to discover configurational cues than quantitative patterns, and no other participant detected quantitative patterns without showing sensitivity to configurational structure, the behavior of SO and SA was very atypical. The data from the other two syntactically impaired participants fit into the range of normal behavior. Behavior from the participant with global aphasia (JB) was affected by Sequence Type at a marginally significant level. The data suggest that he detected configurational relations, namely that all stimuli of one class precede the stimuli from the other class. PR's behavior displayed

To further test the syntactically impaired participants' ability to detect sequential structures, a second AGL experiment was devised. While Experiment 1 used a quantitative rule as a “control pattern”, it was also possible that counting overrode attention to configurational cues. Therefore, Experiment 2 investigated configurational processing only. It used the finite-state grammar A þ B þ , which generates sequences in which a number of stimuli of class A are followed by a number of stimuli of class B (e.g., AAABB, AAABBB, AABBBB). The numbers do not have to match for a sequence to be grammatical. Therefore only the detection of configurational cues was necessary to successfully distinguish grammatical from ungrammatical sequences. It was hypothesized that the absence of a quantitative rule might make it easier to process configuration. Visual stimuli were the same as in Experiment 1. 2.2.1. Participants The syntactically-impaired participants SA, SO, PR and JB were tested. The time between Experiments 1 and 2 was 19 weeks for each participant. 2.2.2. Methods The training phase contained 72 sequences generated by the target grammar A þ B þ . Stimulus blocks of one class could be either two, three or four items long (A2, A3, A4, B2, B3, B4). Each possible block of stimuli of class A was combined with each possible block of class B, resulting in nine different structures (A2B2, A2B3, A2B4, A3B2, etc.). Each structure appeared eight times in the training phase. Thus, in 24 sequences, the numbers of A and B stimuli matched (A2B2, A3B3, A4B4). In 48 sequences, the numbers did not match. Maximum sequence length was eight. The training phase was divided into three blocks of equal length. Participants could take a short break between blocks. The test phase contained 120 novel sequences, half of which were grammatical. There were the same nine different structures of A þ B þ sequences in the grammatical stimuli. Sequences of the types A2B2, A3B3 and A4B4 each appeared eight times. The other grammatical structures appeared six times each. There were five different types of violation sequences: 1. Type 1 (B þ A þ ; e.g., BBAAA) had an illegal initial and final item and contained the illegal bigramnBA. 2. Type 2 ((BA)n; e.g., BABABA) had an illegal initial and final item, contained more than one instance of the illegal bigramnBA. Items of one class did not appear together. 3. Type 3 (initial; e.g., BAAAB) had an illegal initial item, contained the illegal bigramnBA at the initial position of the sequence. Items of one class did not appear together. 4. Type 4 (center; e.g., AABAB) contained the illegal bigramnBA at the center position of the sequence. Items of one class did not appear together. 5. Type 5 (final; e.g., AABBA) had an illegal final item, contained the illegal bigramnBA at the final position of the sequence. Items of one class did not appear together. In Experiment 2, grammatical and ungrammatical test sequences differed substantially in terms of ACS, CN and AS. Grammatical sequences had higher ACS and AS, as well as lower CN. Independent t-tests showed that the difference in ACS was

34


Table 9 Behavioral templates for the analysis of Experiment 2 (numbers represent violation types). Template

Criteria for grammaticality

Ungrammatical sequences according to template

Grammatical sequences according to template

AþBþ (A þ B þ ) þ(B þ A þ ) INIFIN A(A þB)n (A þ B)nB

A number of As is followed by a number of Bs All items of one class precede all items of the other class Sequence starts with A and ends with B Sequence starts with A Sequence ends with B

Viol Viol Viol Viol Viol

AþBþ Viol 1, A þ B þ Viol 4, A þ B þ Viol 4, 5, A þ B þ Viol 3, 4, A þ B þ

Table 10 Rejection patterns across all test phase sequence types for syntactically-impaired participants. Viol. 1: B þ A þ (e.g., BBAAA). Viol. 2: BAn (e.g., BABABA). Viol 3: illegal initial item, violation of block consistency (e.g., BAAAB). Viol 4: illegal central item, violation of block consistency (e.g., AABAB). Viol. 5: illegal final item, violation of block consistency (e.g., AABBA). (Viol. ¼violation type). Sequence type Viol. type AþBþ (grammatical) 1 (B þ A þ )

SA 45.76 SO 60 PR 0 JB 16.67

58.33 75 100 50

Viol. type 2 ([BA]n)

Viol. type 3 (initial)

Viol. type 4 (center)

Viol. type 5 (final)

0 8.33 100 91.67

81.82 58.33 100 100

54.55 58.33 100 50

41.67 75 100 83.33

Table 11 D scores of all eight templates for syntactically-impaired participants. Numbers in bold signify the “best template”, i.e., the highest D score of an individual participant whose behavior was significantly affected by sequence type. A þ B þ : A number of As is followed by a number of Bs. (A þ B þ ) þ(B þ A þ ): All items of one class precede all items of the other class. EQUAL: Equal number of As and Bs. INIFIN: Sequence starts with A and end with B. A(A þ B)n: Sequence starts with A. (A þ B)nB: Sequence ends with B. Template D score

SA SO PR JB

AþBþ

(A þ B þ ) þ (B þ A þ )

EQUAL

INIFIN

A(A þ B)n

(Aþ B)nB

1.61 5 100 58.33

3.44 12.5 83.33 59.03

65.4 88.89 0 6.94

1.49 5.56 83.33 59.03

0.72 14.68 71.43 46.35

17.52 6.75 71.43 39.1

highly significant, t(118) ¼ 6.159, po .001, as were differences in AS, t(118) ¼7.787. p o.001, and CN, t(118) ¼4.435, p o.001. Violation sequences contained two, three or four stimuli of each class A and B. Each violation type occurred 12 times. Appendix C shows the distribution of violation types across different sequences (distinguished by number of A and B stimuli). The test phase was divided into three blocks, with a short break before each block. The stimulus shapes and presentation times in Experiment 2 were the same as in Experiment 1. 2.2.3. Data analysis overview As in Experiment 1, individual data were analyzed via the template analysis. Because of the different grammar, different templates were used (Table 9). 2.2.4. Results Syntactically impaired participants rejected 35.92% (SEM ¼ 12.82) of the grammatical sequences. They rejected 72.92% (SEM¼8.85) of the sequences of violation type 1, 36.17% (SEM¼22.65) of violation type 2, 62.54% (SEM ¼19.48) of violation

1–5 2–5 1–3, 5 1–3 1, 2, 5

type 3, 61.55% (SEM ¼13.95) of violation type 4, and 58.33% (SEM¼18.32) of violation type 5. Logistic regressions found a significant effect of Sequence Type for PR and JB (Appendix D). Individual rejection data are reported in Table 10. It is striking that SA and SO consistently accepted violations of type 2 despite these sequences representing the most flagrant deviations from the target grammar. These violations structured (BA)n had matching numbers of class A and B stimuli, and one possibility is that SA and SO were perseverating on the use of the EQUAL criteria from Experiment 1. We therefore considered the template EQUAL to account for behavior based on counting only. Individual D scores are presented in Table 11. SA and SO consistently rejected sequences in which the number of As and Bs mismatched. Their best template was EQUAL with D scores of 65.4 and 88.89, respectively. PR rejected all violations of the target grammar and accepted all grammatical sequences, achieving a D score of 100 on the A þ B þ template. JB consistently rejected violations of configurational relations. However, his D scores matched on the templates (A þ B þ )þ (B þ A þ ) and INIFIN. Both met above chance criteria. 2.2.5. Discussion Experiment 2 investigated configurational processing in absence of quantitative rules. Two agrammatic participants, SA and SO, displayed behavior that did not match the profiles of participants without syntactic impairment observed in Experiment 1. Both rejected sequences in which the number of stimuli of classes A and B mismatched, but failed to reject violations of configurational relations. This is especially surprising given that during the training phase of Experiment 2 they were exposed to sequences with mismatching numbers. While behavior from Experiment 1 suggests that they understand AGL experiment instructions, it is possible that they assumed that Experiment 2 was a repetition of Experiment 1 and therefore maintained their decision criteria. In future studies, different order of experiments could evaluate this possibility. After showing no evidence of syntactic learning in Experiment 1 according to the template analysis, PR achieved a perfect D score on the template A þ B þ . JB also managed to detect configurational relations, although it is difficult to establish the criteria of his decision due to his D score being the same for two different templates. The behavior of both PR and JB was consistent with the observations made in healthy controls, revealing sensitivity to configuration.

3. General discussion We investigated AGL on an individual level in people with aphasia, both with and without syntactic impairment, as well as healthy controls. In an experiment using the target grammar AnBn, healthy controls as well aphasic participants without syntactic impairment conformed to previously described group (Hochmann et al., 2008; de Vries et al., 2008) and individual (Zimmerer et al.,


35

Table 12 D scores of all eight templates for healthy participants (see also Table 2). Numbers in bold signify the “best template”, i.e. the highest D score of an individual participant that meets the criteria for above chance behavior. AnBn: A number of As is followed by an equal number of Bs. EVEN: Total number of items in a sequence is even. EQUAL: Equal number of As and Bs. A(A þB)n: Sequence starts with A. (A þ B)nB: Sequence ends with B. A þ B þ : All As precede the Bs. (A þ B þ ) þ (B þ A þ ): Items of one class appear together. (AnBn) þ(BnAn): Items of one class appear together, equal number of As and Bs. Template D score

50F 53F1 53F2 58F 60F 62F 66M 67F 73M 78F

AnBn

EVEN

EQUAL

A(A þB)n

(A þ B)nB

AþBþ

(A þ B þ ) þ(B þ A þ )

(AnBn)þ (BnAn)

100 86.44 100 93.1 69.4 66.38 89.92 74.4 2.45 95

55.56 35.75 55.56 38.14 29.52 28.27 55.56 11.67 20.71 54.63

62.5 35.04 62.5 52.59 42 42.28 62.5 28.9 5.55 61.46

62.5 70.83 62.5 63.44 49.06 42.28 47.5 55.11 8.06 61.46

62.5 66.36 62.5 63.44 38.58 35.55 47.5 55.11 7.66 61.46

71.43 76.90 71.43 71.95 50.31 47.14 60.34 66.24 1.23 66.27

62.5 65.44 62.5 62.77 43.82 46.33 63.16 64.89 2.82 56.25

83.33 66.31 83.33 75.84 57.33 58.72 83.33 62.25 1.73 78.47

Table 13 D scores of all eight templates for aphasic participants without syntactic impairment. Numbers in bold signify the “best template”, i.e. the highest D score of an individual participant that meets the criteria for above chance behavior. AnBn: A number of As is followed by an equal number of Bs. EVEN: Total number of items in a sequence is even. EQUAL: Equal number of As and Bs. A(A þ B)n: Sequence starts with A. (A þ B)nB: Sequence ends with B. A þ B þ : All As precede the Bs. (A þ B þ ) þ(B þ A þ ): Items of one class appear together. (AnBn) þ(BnAn): Items of one class appear together, equal number of As and Bs. Template D score

42M 51M 65M1 65M2 84F

AnBn

EVEN

EQUAL

A(Aþ B)n

(A þ B)nB

AþBþ

(A þ B þ ) þ (B þ A þ )

(AnBn)þ (BnAn)

33.4 49.77 76.67 10.14 37.49

8.52 24.45 24.07 0.95 24.38

15.37 17.11 42.71 7.82 27.46

43.26 61.18 58.33 23.12 59.51

51.64 71.62 47.92 9.04 50.83

53.6 72.02 58.73 17.97 66.43

47.73 55.96 47.92 14.66 64.9

21.92 26 60.42 4.48 23.85

Table 14 Distribution of violation types across sequences with different numbers of A and B items in Experiment 2. No. of A items

No. of B items

No. of sequences

2

2

8

2

3

6

2

4

6

3

2

6

3

3

8

3

4

6

4

2

6

4

3

6

4

4

8

Distribution of violation types (n sequences) Type 1 (1), type (1), type 5 (1) Type 1 (2), type (1) Type 1 (2), type (1) Type 1 (1), type (2) Type 1 (1), type (1), type 5 (1) Type 1 (2), type (1) Type 1 (1), type (2) Type 1 (1), type (2) Type 1 (1), type (1), type 5 (1)

2 (4), type 3 (1), type 4 3 (2), type 4 (1), type 5 3 (2), type 4 (2), type 5 3 (1), type 4 (2), type 5 2 (4), type 3 (1), type 4 3 (2), type 4 (1), type 5 3 (1), type 4 (2), type 5 3 (1), type 4 (2), type 5 2 (4), type 3 (1), type 4

2011) profiles. Participants who showed syntactic learning detected either configurational regularities, or configurational as well as quantitative regularities. Two of the syntactically impaired participants, SA and SO, displayed atypical behavior. Despite identifying quantitative patterns, they did not utilize configurational information. In Experiment 2, using the target grammar A þ B þ which is defined only by configuration, they continued to ignore this information even when test sequences contained severe violations. The other

participants classified as syntactically impaired using reversible sentence comprehension tests performed within the normal range on one (PR) or more (JB) AGL tests. After not consistently applying syntactic criteria in Experiment 1, agrammatic participant PR achieved a perfect performance in Experiment 2. Participant JB, classified as globally aphasic, detected configurational information in both experiments. The different patterns of performance within the syntactically impaired group provides some support for multiple processes engaged in syntactic behavior, and potentially different underlying sources of impairment within agrammatism. However, the exact relationship between AGL behavior and performance on conventional tests of grammatical ability is not clear. Whether syntactically impaired patients showed atypical AGL could not be predicted by their performance on conventional tests of grammatical ability. These patients all performed poorly on reversible sentence comprehension tests. However, on grammaticality judgments, the two patients with lower scores showed evidence for sensitivity to configurational information, while the two patients with the better performances showed configurational insensitivity. Dissociation between performance on reversible sentence and grammaticality judgment tasks has been observed previously (Linebarger, Schwartz, & Saffran, 1983). Reversible sentence comprehension requires full processing of configuration of known linguistic forms as well as semantic interpretation. Grammaticality judgments however require only recognition of known configurational relations and forms with no need for further interpretation. AGL probes sensitivity to novel configuration relations expressed in a novel set of forms. The inter-relationship between performances on these different syntactic tasks and the varied demands they make upon processing of lexicalsemantic and novel versus established structural information might be explored in future research.

36


Table 15 Logistic regressions for individual data from Experiments 1 and 2. We report Wald statistics for effects of ACS, CN and sequence type (S. Type; p o.05n, p o .001nn). For the categorical variable sequence type, the model added a rejection if it was empty, and subtracted a rejection if it was full to allow the analysis to converge. Participant

Controls (aged 50þ ) 50F 53F1 53F2 58F 60F 62F 66M 67F 73M 78F

Experiment 1, target grammar: AnBn ACS (Wald)

AS(Wald)

CN (Wald)

S. Type (Wald)

Adjusted cells

2.154 3.483 .003 .269 .530 .086 2.635 .717 .148 .56

.048 .001 .205 .2 .511 1.229 .501 .41 .1773 .088

3.888n 4.411n .592 .33 1.569 1.275 .911 .167 .127 3.482

16.807nn 14.076nn 24.570nn 31.103nn 30.221nn 30.187nn 25.835nn 22.286nn 4.261 26.512nn

Adjusted Adjusted Adjusted Adjusted

.121

16.637nn 24.1nn 33.427nn 2.095 25.146nn

Aphasia without syntactic impairment 42M .088 51M .969 .279 65M1 65M2 .938 84F 1.642

.09 1.533 1.694

1.414 .778 .385 .473 1.659

Aphasia with syntactic impairment SA .174 SO .573 PR 1.638 JB .293

.818 .712 1.160 .365

.665 1.565 3.136 .247

28.865nn 26.554nn 8.963 10.324

Experiment 2, target grammar: A þ B þ SA .051 SO .33 PR .497 JB .182

.657 .012 1.948 .732

.022 .647 4.857 .064

3.580 8.430 14.833nn 16.795n

0

The AGL paradigm has been used to examine mechanisms that underpin grammatical behavior and can contribute to investigations of aphasia. In particular AGL studies may be able to identify impairment of processing configurational/sequence structure information without the requirement of lexical-semantic interpretation. Results may help explain linguistic behavior. For instance, the understanding of English canonical active transitives (The man kills the lion) which have little inflection requires sensitivity to configuration in order to determine the agent (pre-verb NP) and patient (post-verb NP). It is possible, therefore, that impaired sensitivity to configuration revealed by AGL may underlie SA's and SO's agrammatic comprehension performance and in particular their insensitivity to word order information. By contrast, PR and JB's behavior suggests they retain the capacity to detect syntactic regularities despite chance level performance on reversible sentence tests. In the case of JB, severely impaired lexical comprehension may mask residual syntactic processing ability and, while tasks involving natural language forms are unable to detect this retained capacity, AGL tasks may be particularly well suited to explore sequential processing ability in the face of global aphasia. The proposed explanation leads to two predictions that need to be explored in subsequent investigations of patient populations: First, individuals who display the abnormal profile reported here are likely to have difficulties with natural language syntax, including comprehension of sentences reliant on word order such as English canonical actives. Second, individuals who have no difficulties with sentences reliant on word order will display an AGL profile within the normative range of performance. While the current study was designed to investigate a wider variety of different behaviors (templates), future work may aim to limit hypotheses to three different types of behavior: (1) configurational only, (2) quantitative only, and (3) configurational and quantitative.

Adjusted

Adjusted

Adjusted Adjusted

Adjusted Adjusted Adjusted

Adjusted

Christiansen et al. (2010), discussing the evidence from AGL in aphasia, suggest that there are multiple inter-related mechanisms mediating sequential behavior across cognitive domains. Indeed, the evidence of dissociation between agrammatism and impairments of other higher cognitive capacities supports this view. For example, people with severe agrammatic aphasia remain able to process the structural properties of mathematical equations (Klessinger, Szczerbinski, & Varley, 2007; Varley, Klessinger, Romanowski, & Siegal, 2005) and also display sequential organized behavior in a range of other domains like navigation (Bek, Blades, Siegal, & Varley, 2010), communicative problem-solving (Willems, Benn, Hagoort, Toni, & Varley, 2011) and theory of mind (Apperly, Samson, Carroll, Hussain, & Humphreys, 2006; Varley, Siegal, & Want, 2001). It is possible that different cognitive domains draw upon different subsets of mechanisms, with varying degrees of overlap. Resources required for learning of new configurational relations in AGL may be essential for linguistic syntactic behavior, but are not necessarily employed in other established cognitive domains such as math or theory of mind reasoning. More data are needed to establish the relationship between AGL, syntax and other higher cognitive functions. The interpretation of AGL learning patterns in neurological patients is complicated by the observation that some healthy controls fail to learn syntactic properties of stimulus sequences (see also Visser et al., 2009; Zimmerer et al., 2011). Patients who show no evidence of learning in AGL can therefore not be considered atypical when compared against control behavior. One key issue is the refinement of AGL methods. In particular, passive observation during training (employed by the majority of AGL studies) may result in some participants failing to recruit attentional resources during training, resulting in chance behavior in the test phase. One possibility is that grammatical variations of serial reaction time tasks may elicit more consistent measurements of syntactic behavior


(Hunt & Aslin, 2010; Misyak et al., 2010; Zimmerer, Indefrey, & Saddy, submitted). Serial reaction time tasks involve a mapping between each presented stimulus and a response button. After stimulus onset, participants have to press the corresponding button as quickly as possible. Sequences of stimuli are structured. If a perturbation to the sequence is introduced, participants who have learned properties of the structure display slower response times due to mismatch between expectation and event. Conventional AGL methodology also lends itself to explicit strategizing. Studies that have explored post-experimental awareness of decision criteria have found that participant reports match the behavior observed in the test phase (Dulany, Carlson, & Dewey, 1984; Perruchet & Pacteau, 1990; Zimmerer et al., 2011). Again, use of reaction time tests may provide an elegant solution for the problem of conscious strategies since the manual task diverts attention from the regularities underlying sequence structure. While these concerns require continued investigation and development of AGL methodology, the observation that some individuals with agrammatism show profoundly different syntactic behavior even in absence of phonological and lexical-semantic information remains. AGL may represent a useful clinical tool for diagnosis of different sub-types of agrammatic impairment. As a diagnostic tool, AGL might distinguish between patients with intact versus impaired basic structure processing systems. Furthermore, where atypical AGL profiles are identified, AGL training might represent a valuable precursor step in intervention. Hoen et al. (2003) trained agrammatic participants on a transformational grammar (see Section 1) using a card order manipulation task and reported improvement in some sentence structures. A clear advantage of AGL is that it can employ grammars much simpler than those of natural language, and increase complexity with time. AGL intervention early after onset may support disrupted syntactic processing systems and keep them latent for subsequent interventions with sentences.

Acknowledgements We would like to thank all who took part in the experiments, as well as Lucy Dyson for her help with recruitment and data collection. We thank Ruth Herbert and James Douglas Saddy for their comments on the project, and Philip T. Smith for his advice on data analysis. This article is partially based on a PhD dissertation by Zimmerer (2010). The research was partially supported by the ESRC grant RES-051-27-0189.

Appendix A See Table 12.

Appendix B See Table 13.

Appendix C See Table 14.

Appendix D See Table 15.

37

References Ankerstein, C. A., Peace, V., Evans, J., & Varley, R. Digit span in recall and recognition, unpublished. Apperly, I. A., Samson, D., Carroll, N., Hussain, S., & Humphreys, G. (2006). Intact first- and second-order false beliefe reasoning in a patient with severely impaired grammar. Social Neuroscience, 1(3–4), 334–348. Bahlmann, J., Schubotz, R. I., & Friederici, A. D. (2008). Hierarchical artificial grammar processing engages Broca's area. NeuroImage, 42, 525–534. Bahlmann, J., Schubotz, R. I., Mueller, D., Koester, A. D., & Friederici, A. D. (2009). Neural circuits of hierarchical visuo-spatial sequence processing. Brain Research, 1298, 161–170. Bek, J., Blades, M., Siegal, M., & Varley, R. (2010). Language and spatial reorientation: Evidence from severe aphasia. Journal of Experimental Psychology: Learning, Memory, and Cognition, 36(3), 646–658. Berndt, R. S., Mitchum, C. C., & Haendiges, A. N. (1996). Comprehension of reversible sentences in “agrammatism”: A meta-analysis. Cognition, 58(3), 289–308. Berndt, R. S., & Caramazza, A. (1999). How "regular" is sentence comprehension in Broca's aphasia? It depends on how you select the patients. Brain and Language, 67(3), 242–247. Biassou, N., Obler, L. K., Nespoulous, J.-L., Dordain, M., & Harris, K. S. (1997). Dual processing of open- and closed-class Words. Brain and Language, 57(3), 360–373. Caplan, D., Baker, C., & Dehaut, F. (1985). Syntactic determinants of sentence comprehension in aphasia. Cognition, 21, 117–175. Caplan, D., & Waters, G. S. (1999). Verbal working memory and sentence comprehension. Behavioral and Brain Sciences, 22, 77–126. Caplan, D., Waters, G. S., DeDe, G., Michaud, J., & Reddy, A. (2007). A study of syntactic processing in aphasia I: Behavioral (psycholinguistic) aspects. Brain and Language, 101(2), 103–150. Caramazza, A., Capitani, E., Rey, A., & Berndt, R. S. (2001). Agrammatic Broca's aphasia is not associated with a single pattern of comprehension performance. Brain and Language, 76(2), 158–184. Caramazza, A., & Zurif, E. B. (1976). Dissociation of algorithmic and heuristic processes in language comprehension: Evidence from aphasia. Brain and Language, 3, 572–582. Chomsky, N. (1957). Syntactic structures. Mouton: The Hague. Christiansen, M. H., Louise Kelly, M., Shillcock, R. C., & Greenfield, K. (2010). Impaired artificial grammar learning in agrammatism. Cognition, 116(3), 382–393. Conway, C. M., & Pisoni, D. B. (2008). Neurocognitive basis of implicit learning of sequential structure and its relation to language processing. Annals of the New York Academy of Sciences, 1145, 113–131. de Vries, M. H., Monaghan, P., Knecht, S., & Zwitserlood, P. (2008). Syntactic structure and artificial grammar learning: The learnability of embedded hierarchical structures. Cognition, 107(2), 763–774. Della Sala, S., Gray, C., Baddeley, A., & Wilson, L. (1997). The visual patterns test: A new test of short-tern visual recall. Suffolk, U.K.: Thames Valley Test Company. Dominey, P. F., Hoen, M., Blanc, J. -M., & Lelekov-Boissard, T. (2003). Neurological basis of language and sequential cognition: Evidence from simulation, aphasia, and ERP studies. Brain and Language, 86(2), 207–225. Drai, D., & Grodzinsky, Y. (2006). A new empirical angle on the variability debate: Quantitative neurosyntactic analyses of a large data set from Broca's aphasia. Brain and Language, 96(2), 117–128. Druks, J. (2002). Verbs and nouns: A review of the literature. Journal of Neurolinguistics, 15(3–5), 289–315. Dulany, D. E., Carlson, R. A., & Dewey, G. I. (1984). A case of syntactical learning and judgment: How conscious and how abstract? Journal of Experimental Psychology: General, 113(4), 541–555. Fitch, W. T., & Friederici, A. (2012). Artificial grammar learning meets formal language theory: An overview. Philosophical Transactions of the Royal Society B, 367, 1933–1955. Fitch, W. T., & Hauser, M. D. (2004). Computational constraints on syntactic processing in a nonhuman primate. Science, 303, 377–380. Forster, K. I., & Forster, J. C. (2003). DMDX: A windows display program with millisecond accuracy. Behavior Research Methods, Instruments & Computers, 35 (1), 116–124. Franklin, S., Turner, J. E., & Ellis, A. W. (1992). The ADA auditory comprehension battery. York, U.K.: University of York. Friederici, A. D. (1982). Syntactic and semantic processes in aphasic deficits: The availability of prepositions. Brain and Language, 15(2), 249–258. Friederici, A. D., Bahlmann, J., Heim, S., Schubotz, R. I., & Anwander, A. (2006). The brain differentiates human and non-human grammars: Functional localization and structural connectivity. Proceedings of the National Academy of Sciences, 103 (7), 2458–2463. Grodzinsky, Y. (2000). The neurology of syntax: Language use without Broca's area. Behavioral and Brain Sciences, 23, 1–71. Haarmann, H. J., Just, M. A., & Carpenter, P. A. (1997). Aphasic sentence comprehension as a resource Deficit: A computational approach. Brain and Language, 59(1), 76–120. Haarmann, H. J., & Kolk, H. H. J. (1994). On-line sensitivity to subject-verb agreement violations in Broca's aphasics: The role of syntactic complexity and time. Brain and Language, 46(4), 493–516.

38


Hauser, M. D., Barner, D., & O'Donnel, T. (2007). Evolutionary linguistics: A new look at an old landscape. Language Learning and Development, 3(2), 101–132. Hauser, M. D., Chomsky, N., & Fitch, W. T. (2002). The faculty of language: What is it, who has it, and how did it evolve? Science, 298, 1569–1579. Hochmann, J. R., Azadpour, M., & Mehler, J. (2008). Do humans really learn AnBn artificial grammars from exemplars? Cognitive Science, 32(6), 1021–1036. Hoen, M., Golembiowski, M., Guyot, E., Deprez, V., Caplan, D., & Dominey, P. F. (2003). Training with cognitive sequences improves syntactic comprehension in agrammatic aphasics. Neuroreport, 14(3), 495–499. Hunt, R. H., & Aslin, R. N. (2010). Category induction via distributional analysis: Evidence from a serial reaction time task. Journal of Memory and Language, 62, 98–112. Ishihara, S. (2007). Ishiharas test chart for color deficiency. 14 Plate edition. Tokyo, Japan: Kanehara Trading Inc. Kay, J., Lesser, R., & Coltheart, M. (1992). Psycholinguistic assessment of language processing in aphasia. Hove, U.K.: Psychology Press. Klessinger, N., Szczerbinski, M., & Varley, R. (2007). Algebra in a man with severe aphasia. Neuropsychologia, 45(8), 1642–1648. Knowlton, B. J., & Squire, L. R. (1994). The information acquired during artificial grammar learning. Journal of Experimental Psychology: Learning, Memory and Cognition, 20(1), 79–91. Linebarger, M. C., Schwartz, M. F., & Saffran, E. M. (1983). Sensitivity to grammatical structure in so-called agrammatic aphasics. Cognition, 13(3), 361–392. Marcus, G. F., Vijayan, S., Bandi Rao, S., & Vishton, P. M. (1999). Rule learning by seven-month-old infants. Science, 283(5398), 77–80. Martin, R. C. (2006). The neuropsychology of sentence processing: Where do we stand? Cognitive Neuropsychology, 23(1), 74–95. Misyak, J. B., & Christiansen, M. H. (2012). Statistical learning and language: An individual differences study. Language Learning, 62(1), 302–331. Misyak, J. B., Christiansen, M. H., & Tomblin, J. B. (2010). Sequential expectations: The role prediction-based learning in language. Topics in Cognitive Science, 2, 138–153. Miyake, A., Carpenter, P. A., & Just, M. A. (1994). A capacity approach to syntactic comprehension disorders: Making normal adults perform like aphasic patients. Cognitive Neuropsychology, 11(6), 671–717. Perruchet, P., & Pacteau, C. (1990). Synthetic grammar learning: Implicit rule abstraction or explicit fragmentary knowledge? Journal of Experimental Psychology: General, 119, 264–275. Perruchet, P., & Rey, A. (2005). Does the mastery of center-embedded linguistic structures distinguish humans from nonhuman primates? Psychonomic Bulletin & Review, 12(2), 307–313.

Petersson, K. M., Folia, V., & Hagoort, P. (2012). What artificial grammar learning reveals about the neurobiology of syntax. Brain and Language, 120(2), 83–95. Petersson, K. M., Forkstam, C., & Ingvar, M. (2004). Artificial syntactic violations activate Broca's region. Cognitive Science, 28(3), 383–407. Pinker, S., & Jackendoff, R. (2005). The faculty of language: What's special about it? Cognition, 95(2), 201–236. Pothos, E. M. (2007). Theories of artificial grammar learning. Psychological Bulletin, 133(2), 227–244. Pujol, J., Deus, J., Losilla, J. M., & Capdevila, A. (1999). Cerebral lateralization of language in normal left-handed people studied by functional MRI. Neurology, 53(5), 1038–1043. Reber, A. S. (1967). Implicit learning of artificial grammars. Journal of Verbal Learning & Verbal Behavior, 6(6), 855–863. Redington, M., & Chater, N. (1996). Transfer in artificial grammar learning: A reevaluation. Journal of Experimental Psychology: General, 125(2), 123–138. Saffran, E. M., Schwartz, M. F., & Linebarger, M. C. (1998). Semantic influences on thematic role assignment: Evidence from normals and aphasics. Brain and Language, 62(2), 255–297. Saffran, J. R., Hauser, M. D., Seibel, R., Kapfhamer, J., Tsao, F., & Cushman, F. (2008). Grammatical pattern learning by human infants and cotton-top tamarin monkeys. Cognition, 107, 479–500. Varley, R. A., Klessinger, N. J. C., Romanowski, C. A. J., & Siegal, M. (2005). Agrammatic but numerate. Proceedings of the National Academy of Sciences of the United States of America, 102(9), 3519–3524. Varley, R. A., Siegal, M., & Want, S. C. (2001). Severe impairment in grammar does not preclude theory of mind. Neurocase, 2001(7), 489–493. Visser, I., Raijmakers, M. E. J., & Pothos, E. M. (2009). Individual strategies in artificial grammar learning. American Journal of Psychology, 122(3), 193–307. Wassenaar, M., & Hagoort, P. (2007). Thematic role assignment in patients with Broca's aphasia: Sentence-picture matching electrified. Neuropsychologia, 45(4), 716–740. Willems, R. M., Benn, Y., Hagoort, P., Toni, I., & Varley, R. A. (2011). Communication without a functioning language system. Neuropsychologia, 49, 3130–3135. Zimmerer, V. C. (2010). Intact and impaired fundamentals of syntax: Artificial grammar learning in healthy speakers and people with aphasia. University of Sheffield (Ph.D. Dissertation). Zimmerer, V. C., Cowell, P. E., & Varley, R. A. (2011). Individual behavior in learning of an artificial grammar. Memory & Cognition, 39(3), 491–501. Zimmerer, V. C., Indefrey, P., & Saddy, J. D. Implicit syntactic learning in a reaction time task, submitted.

The P600 in Implicit Artificial Grammar Learning.

Auditory artificial grammar learning in macaque and marmoset monkeys.

Implicit learning and implicit treatment outcomes in individuals with aphasia.

Metrical presentation boosts implicit learning of artificial grammar.

Mixed-complexity artificial grammar learning in humans and macaque monkeys: evaluating learning strategies.

EEG potentials associated with artificial grammar learning in the primate brain.

Implicit learning and reading: insights from typical children and children with developmental dyslexia using the artificial grammar learning (AGL) paradigm.

Budgerigars and zebra finches differ in how they generalize in an artificial grammar learning experiment.

Artificial grammar learning in zebra finches and human adults: XYX versus XXY.

Concurrence of rule- and similarity-based mechanisms in artificial grammar learning.

Learning of grammar-like visual sequences by adults with and without language-learning disabilities.

Discovery of a Recursive Principle: An Artificial Grammar Investigation of Human Learning of a Counting Recursion Language.

Different Cognitive Profiles of Patients with Severe Aphasia.

Hidden word learning capacity through orthography in aphasia.

Story grammar: effective literature instruction for high school students with learning disabilities.

Neuropathologic Associations of Learning and Memory in Primary Progressive Aphasia.

Using mobile technology with individuals with aphasia: native iPad features and everyday apps.

Semantic Knowledge Use in Discourse Produced by Individuals with Anomic Aphasia.

Effects of intensive phonomotor treatment on reading in eight individuals with aphasia and phonological alexia.

Quantifying the effort individuals with aphasia invest in working memory tasks through heart rate variability.

A Novel Pupillometric Method for Indexing Word Difficulty in Individuals With and Without Aphasia.

Reliability of DISCUS rating in individuals with learning disabilities.

Test-enhanced learning versus errorless learning in aphasia rehabilitation: testing competing psychological principles.

Studies of exclusion in individuals with severe mental retardation.