Presentation on theme: "Background Adults and infants appear to be able to quickly learn statistical properties of language input (Saffran, et al.,1997; Saffran, 1996; Maye, Werker,"— Presentation transcript:
Background Adults and infants appear to be able to quickly learn statistical properties of language input (Saffran, et al.,1997; Saffran, 1996; Maye, Werker, & Gerken, 2002). Statistical properties of language have been shown to be potentially useful in providing infants with early information for segmentation of phonemes and words. Statistical learning has also been shown to potentially useful for the more difficult nonadjacent relations necessary for syntax (Gómez, 2002; Newport & Aslin, 2004) Statistical learning has been proposed as a basic biological trait that enables language learning (Christiansen & Ellefson, 2002; Tomesello, 2003). Implicit Learning in Adolescents with and without Specific Language Impairment J. Bruce Tomblin and Xuyang Zhang The University of Iowa, Iowa City Discussion This study asked whether the statistical learning rates of children with poor language skills as represented in adolescents with SLI was slower than in normal language age mates. Statistical learning was implemented using a serial response task that requires speeded manual responses. The presence of a predictable sequence resulted in systematically faster responses for all participants. Language status was associated with generally slower responses and with differential rate of learning as reflected in a significant quadratic component being associated with better language. Thus, children with better language learn the statistical dependencies of the sequences more quickly than the children with poorer language skill. Figure 2 shows the characteristics of learning by plotting the growth parameters for children with low language and those with high language abilities. The high language skill children improve quickly and then appear to reach an asymptote. The low language skill children progress more regularly across trials. The results suggest that basic statistical learning abilities may contribute to the language learning characteristics of children with SLI. Figure 1a (left panel) displays the median response times in each block for the controls and the SLI. The plotted values are shown in the accompanying Table. Language items used for computation of language scores. Figure 1b shows the same performance plotted by 10 trial set across the blocks. Abstract Implicit learning of a visual sequence was examined in adolescents with and without specific language impairment using a serial response task. Participants were presented with an array of 4 boxes in which an object could appear. They were to push a button associated with the box as soon as they saw the object. Blocks of trials were arranged such that the object either appeared randomly or according to a pattern. Response times for correct trials showed that responses for both groups improved in the trial blocks containing the pattern. Adolescents with SLI were slower across all conditions, but also showed slower learning rates during the pattern learning in comparison to the controls. The results suggest that mechanisms involved with implicit learning may account for some of the language learning problems of individuals with SLI. Acknowledgement This study was supported by contract NIH-DC from the National Institute on Deafness and Other Communication Disorders and clinical research center grant P0-DC-02748, also from the National Institute on Deafness and Other Communication Disorders. Language development reflects adaptive changes to complex network(s) in response to linguistic and nonlinguistic exposure. The nature of these networks in response to stimuli can be seen in short-term learning tasks. Thus, learning performance in a statistical learning task such as a serial response task should be associated with language status. Serial Reaction Time Paradigm Learner identifies the occurrence of each element of a sequentially random or sequentially structured sequence of stimuli (Nissen & Bullemer, 1987). Sequences are blocked with respect to whether they are random or conform to some form of pattern based on a rule based sequence or repeating sequence. Sequence (statistical) learning is reflected in decreased response times (RTs) during the pattern learning and increased response times in random blocks following pattern blocks. Learning in this case is demonstrated via implicit response. Question Do individuals with SLI show slower rates of learning in an SRT task than age matched controls? Hypothesis Methods Participants 43 adolescents with SLI 40 adolescents with normal language Both groups matched on Performance IQ and chronological age GroupMean Age (SD) Mean Perf. IQ (SD) Mean Lang. Composite z-score (SD) SLI14.76 (0.58) (8.1) (0.36) Contro l (0.49) (5.15) -.09 (.81) Clinical Measures Diagnosis was based on measures taken in second grade and confirmed in eighth grade. Language Diagnosis 2 language areas out of 5 below the 10 th percentile Language Diagnosis based on:PPVT-R; CREVT-Expressive Voc.; CELF-III: Recalling Sentences, Sentence Structure; Concepts and Directions, Word Structure (2 nd grd.) Formulated Sentences (8 th grd), Listening to Paragraphs, narrative production. Nonverbal IQ WISC-R All participants had performance IQ >85 RandomPatternRandom 100 trials Serial Response Task A serial response task based on Thomas & Nelson (2001) was used. Control of stimulus presentation and response recording was done via E-prime. Participants were instructed that they would see a set of four boxes. A green monster would appear in one of the boxes. Their job was to press one of four buttons on a button box as soon as the monster appeared. The buttons were arranged in the same order as the boxes and they were to press the button that went with the box. A stimulus trial consisted of one occurrence of the monster. Participants were NOT told that there was any pattern to the sequences. Brief rest break Stimulus trials were organized into blocks of 100. The trials during the first block were randomized. The next two blocks followed a pattern: 2, 4, 1, 3, 4, 2, 1, 4, 3, 1, 2, 4, 1, 3, 4, 2, 1, 4, 3, 1 ….. Within each sequence 1 and 4 occurred less frequently within each sequence (3X) than 2 and 3 (4X) Analysis Reaction times were grouped into sets of 10 trials each and a median reaction time was computed for each set. The association of language ability with rate of statistical learning was compared by use of growth curve analysis using a mixture model (Proc Mixed in SAS). Linear and quadratic properties of RT growth during the pattern blocks were modelled as random effects. The interaction of these growth parameters with the fixed effect of the composite language score provided a test of the relationship between language and statistical learning. RandomPattern Random SLI Contr l Mixed Model ParameterCoefficient p Intercept Linear Slope Quadratic Language Linear*Lang Quad*Lang The test for differences in the rate of learning during the trial sets comprising the pattern condition was performed using Proc Mixed. The results of this analysis are shown in Table 2. The test for differential learning rates involve the fixed effect tests shown in red. The first of these (Language) tests whether there was an overall difference in intercept as a function of language ability. Support for this effect was shown (p=0.001). The second test concerned the extent to which linear changes in RT were associated with language levels. No support for this effect was found (p=.56). The third growth parameter consisted of association of the quadratic component of learning and language level. Support for this effect was found (p=.04). Table 2. Results of learning curve analysis using Proc Mixed testing for parameters of intercept, slope, & quadratic and the relationship of composite language with these parameters. Fixed effects tests are shown in red. Although statistical learning appears important for language learning it is less clear whether there are individual differences in statistical learning ability and thus if variation in such a trait can account for individual differences in language development especially those comprising SLI. Individual Differences in Statistical Learning Reber (1993) has proposed that implicit learning of which statistical learning is a special case, is largely invariant across individuals. A part of this invariance was even found across developmental level. Evidence of developmental invariance has been shown by Howard & Howard (1992) Meulemans et al. (1998). Also implicit learning appears to be preserved in several forms of progressive brain disease such as Alzheimers disease (Reber et al., in press) amnesia (Knowlton & Squire, 1996), closed-head injury (McDowall & Martin, 1996). Evidence for individual differences in implicit learning can also be found. Fletcher et al. (2000) examined children of varying ages and mental abilities and found individual differences in SRT learning were associated with mental age. Vicari et al. (2003) discovered that children and adolescents with dyslexia were less capable of statistical learning in a nonverbal SRT task. Cherry and Stadler (1995) found Individual differences on the SRT task have also been shown to correlate with variations in educational attainment, occupational status, and verbal ability in older adults. Results Median response times for each block within each group (SLI and Control) are shown in Figure 1a. Figure 1b. shows the same responses plotted by set. The two groups were similar during the first random condition. As expected the children with SLI were slower than the controls (Kail, 1994). During the two pattern conditions, each group demonstrated systematic decreases in response times. Also during this time, It can be seen that greater differences emerge for the two groups suggesting that the children with SLI were learning more slowly. Finally, each group slowed substantially when the second random block was presented. In fact, their performance was slower than during the initial random block suggesting an interference from the pattern learning. The data in Figure 1b show that within each group there was considerable variability.