Grammatical Difficulties in Children with Specific Language Impairment: Is Learning Deficient?

Pattern Learning of Grammatical Relations: Evidence from Artificial Grammar Learning

Statistical learning of grammatical relations has been studied using artificial languages in which statistical relationships such as frequency of co-occurrence between units or transitional probabilities are manipulated to give rise to structural dependencies in the language. Gómez and Gerken [1999] showed that 12-month-old infants can track frequency of co-occurrence to learn the orderings of words in sequences. They exposed infants to 1 of 2 artificial languages. Some transitions between word categories occurred in language 1 but never occurred in language 2, and vice versa. After brief exposure to a subset of strings in their training language, infants were tested with novel strings (i.e., strings they had never encountered during training) from both languages. They found that infants listened longer to novel strings from their training language than to strings from the other language, regardless of which language they heard during training. Because infants were never tested on the exact strings encountered during training, it could be concluded that learning was not restricted to memory for particular strings, but rather generalized to novel strings with familiar co-occurrence patterns.

A key difficulty posed by grammar, however, is that it does not just involve learning dependencies between adjacent words. There are also long-distance dependencies. For instance, in ‘The chicken on the ball is black’, it is the chicken, rather than the ball, that is black. Saffran [2001] examined whether statistical learning mechanisms can succeed in mastering more complex structures that are not tied to the surface properties of the input, such as hierarchical phrase structures. For instance, the words in ‘The space probe sent back photographs of Mars’ fall into particular groupings [(The (space probe)) (sent back (photographs of Mars))] rather than random grouping [e.g., (The space) (probe sent back) (photographs of) (Mars)]. To examine if statistical learning could extract hierarchical phrase structure, Saffran [2001] compared learning of 2 artificial languages, 1 containing predictive dependencies between words and the other lacking predictive dependencies. The predictive dependencies were defined as relatively higher transitional probabilities between items. In the predictive language, the presence of a word token in a category was always preceded by an occurrence of a word belonging to a different word category. In contrast, in the nonpredictive language, the relationships between categories were variable. Category membership in the languages can be learned by paying attention to the distribution of words. This is similar to the case that speakers of natural languages can infer the category memberships of novel words from surrounding words, for example, a word preceded by ‘the’ is very likely to be a noun. The training involved a 30-min session on each of 2 consecutive days for adult participants and a 21- to 28-min session for children participants. Both adult and child learners exposed to the language containing predictive dependencies performed better in detecting phrasal units than learners exposed to the language lacking predictive dependencies. Of course, the success in learning does not necessarily entail that the acquired knowledge was hierarchical in nature; nevertheless, the study demonstrates that this type of learning can extend to relationships between word categories that are hierarchically organized, and is not restricted to learning relationships between individual word items, or relationships that can only be sequentially characterized.

Distributional regularities have been shown to be useful in learning other aspects of grammatical relations. These include building syntactic categories from distributional cues in speech [e.g., Gómez & Lakusta, 2004], and learning dependencies between syntactic categories [e.g., Gerken, Wilson, & Lewis, 2005].

Artificial grammar learning has also been used to explore the conditions that promote learning of statistical structure. In order to correctly mark grammatical agreement one typically has to ignore considerable variation in intervening elements – typically open-class words that have a large set size. Consider, for instance, the variable material that can intervene between 2 grammatical elements that need to be marked for agreement in English structures such as progressives (The cat is eating), perfectives (He has played the computer game for hours), plurals (The divers on the boat are excited), and third person singular -s (She jumps). Gómez [2002] considered how variability of intervening elements affected learning of nonadjacent dependencies, using artificial 3-word strings, A-X-B, where A and B were always the same, and X was represented by a set of 3, 12, or 24 words. She exposed 18-monthold infants to 3-word strings composed of nonsense words such as pel-chila-tud, votchila-jic, in which the occurrence of the first word always predicted the identity of the third word. During a subsequent test phase, the infants’ listening time was measured to ‘grammatical’ strings conforming to the dependencies between A and B that had been heard during training and other ‘ungrammatical’ strings that violated these dependencies. Gómez found that infants discriminated between grammatical and ungrammatical strings only in the high variability condition (24 words). She concluded that listeners seek out statistical regularity in the input: if there are strong dependencies between adjacent elements (as with the small set size condition), the infant will focus on these, and less attention is paid to nonadjacent dependencies. However, when the set of intervening words is large, there is little predictability between adjacent words, and therefore the more stable relationship between nonadjacent elements becomes more salient and thus easy to detect. Note that infants in this study demonstrated sensitivity to statistical regularities in the input in the absence of any corrective feedback. In the same report, Gómez demonstrated similar results with adults, who were required to give grammaticality judgments on word strings after exposure to novel word sequences: once again, detection of nonadjacent dependencies was best when the transitional probabilities between adjacent items were low. Similar findings were replicated by Onnis, Christiansen, Chater, and Gómez [2003] and Onnis, Monaghan, Christiansen, and Chater [2004].

Evidence for Exemplar-Based Learning by Children with SLI

Gopnik and Crago [1991] were among the first to suggest that children with SLI engaged in exemplar-based learning. They argued that regular and irregular past tenses were treated the same: the child learned the whole inflected form and did not show awareness that the -ed ending was common to the regulars. Goad and Rebellati [1994] provided evidence that even when children appeared to be aware of a rule, for example plural -s, they did not internalize it, but applied it effortfully by remembering the explicit rule ‘add -s’. These authors were working in a Chomskyan frame-work, and so interpreted their results as indicating a lack of innate grammatical rules; however, the findings could be reconceptualized as evidence for a failure of statistical learning mechanisms.

A tendency to use rote-learned forms has also been reported in studies of SLI focusing on learning syntactic structures of new words. Jones and Conti-Ramsden [1997] compared the verb use of 3 children with SLI and that of their younger siblings matched on mean length of utterance. The children with SLI tended to use lexical verbs in a narrower range of forms than their younger siblings, although their verb lexicons were not dissimilar in size. Interestingly, overlap of utterances with those of the caregiver was greater for children with SLI than for their siblings, suggesting rote learning of syntactic forms. Similar results were reported by Stokes and Fletcher [2000] who investigated spontaneous speech of Cantonese-speaking children with SLI. They found that these children’s use of aspect markers was far less productive than that of language-matched typically developing children. Furthermore, Skipp, Windfuhr, and Conti-Ramsden [2002] reported context-dependent use of newly acquired noun phrases in children with SLI. In their study, a group of 35-month-old children with SLI learned 4 nouns presented in 1 of 4 argument structures: (1) neither agent nor patient (‘Look - Gabber!’), (2) agent only (‘The Mogo is pushing’), (3) patient only (‘kissing the Neffy’), or (4) both agent and patient (‘Minnie is washing the Toma’). They found that typically developing children used the noun words in spontaneous utterances equally in all 4 argument structures, regardless of the argument structures in which the words were presented during training. However, children with SLI demonstrated greater input dependence in terms of the type of arguments they used in spontaneous utterances after training.

Finally, more indirect evidence for difficulties in learning grammatical abstractions comes from consideration of children’s performance on grammatical comprehension tasks, where there is no demand for syntactic creativity. At school age, most children with SLI are able to respond correctly to many syntactically simple sentences in a multiple-choice context. They do not behave as if they have no knowledge of how grammar conveys meaning: rather, they perform inconsistently on more complex constructions such as passives, comparatives, or spatial prepositions [Bishop, 1982]. This was demonstrated in an intervention study by Bishop, Adams, and Rosen [2006], who showed that, even after daily training with a particular construction type, fluent automatic comprehension was not achieved, although performance was well above chance. This performance pattern suggests that these children may be poor statistical learners who have not reached a stage at which sentence comprehension is supported by more abstract syntactic patterns required for consistent and accurate performance.

Artificial Grammar Learning in Language-Impaired Individuals

Learning grammar involves learning both adjacent and nonadjacent relationships. For instance, in English serial word order is important to language comprehension as changing the order of noun phrases in a simple Noun-Verb-Noun sentence (e.g., ‘The boy kicked the girl’ vs. ‘The girl kicked the boy’) would result in meaning differences. Plante, Gómez, and Gerken [2002] examined statistical learning of sequential word order in adults with and without language/learning disabilities. Participants first listened to a set of sentence strings conforming to the word order constraints in a finite-state grammar and then provided grammaticality judgments on novel strings. With only 5-min exposure to the language, typically developing adults were able to exceed chance performance, whereas adults with language/learning disabilities showed chance level performance. Because the speed of presentation of the training stimuli was controlled to be approximately 1.32 words/s, poor performance in the adults with language/learning disabilities, as argued by Plante et al. [2002], was unlikely due to difficulties in processing rapidly presented acoustic information.

Learning of nonadjacent relationships is of particular relevance to the grammatical morpheme deficits that are seen in young typically developing children and that also constitute one of the most salient features in children with SLI. Before full mastery of verb inflections, young English-speaking children go through a phase where they are inconsistent in the use of verb inflections, and produce utterances such as ‘John go there’, while also producing correctly inflected versions of the same verb, for example, ‘He goes there’. Wexler [1994] argued that this represents a stage where children fail to understand that finite verb forms (forms involving tense and agreement) are obligatory in main clauses, and so treat marking of finiteness as optional. This account has been extended to explain deficits in children with SLI by arguing that they have an unusually prolonged optional infinitive stage [Rice et al., 1995]. However, recent empirical studies on pattern learning of remote dependencies in typically developing individuals provide an alternative account.

Grunow, Spaulding, Gómez, and Plante [2006] adopted the task developed by Gómez [2002] to investigate learning of nonadjacent dependencies in college students with and without language-based learning disabilities. Participants in the study listened to ‘sentences’ composed of 3 nonsense words in 1 of 2 variability conditions: the set size of the middle words was either 12 or 24 words. Adults with normal language were able to learn and generalize the nonadjacent dependencies when variability was high (set size 24 words). In contrast, adults with language-based learning disabilities did not perform above chance under either variability condition. Grunow et al. [2006] concluded that the adults with language-based learning disabilities have poor sensitivity to statistical information in speech input. Unfortunately, the results were not watertight, because the sample size was small (n = 11 per condition) and there were no significant group differences in learning nonadjacent dependencies, for either trained items or generalization items. The conclusion therefore hinged just on the contrasting patterns of mastery across groups and conditions. However, confidence in the conclusions is boosted by findings from another study adopting a similar procedure in adolescents with and without language impairment [Hsu, Tomblin, & Christiansen, 2008]. Again, high variability facilitated learning of nonadjacent dependencies in typically developing adolescents but not adolescents with language impairments. When the participants’ language skills were taken into account, a significant, albeit modest, correlation (r = 0.35) in the high-variability condition was found.

Hsu et al. [2008] further explored individual differences in this kind of learning by considering the number of participants who reached 100% accuracy in learning at least 1 nonadjacent pair. Because the test items in their study were not novel strings but strings heard during training, the results would provide further information about exemplar-based learning. Of particular interest is the children’s performance in the variability condition where there were only 2 possible intervening words. In this ‘set size = 2’ condition, the number of different sentence strings is much lower (i.e., 6) and the token frequency is much higher (token frequency = 72 for each sentence) compared to the other 2 variability conditions: in the ‘set size = 12’ condition, there were 36 different sentences and each had a token frequency of 12, and in the ‘set size = 24’ condition, there were 72 different sentences and each had a token frequency of only 6. If there was a sign of exemplar-based learning, one would most likely see such learning in the condition where token frequency is high. The results showed that the proportion of typically developing participants who mastered at least 1 pair was highest in the high-variability condition, as expected. However, for the language-impaired group, 15% mastered at least 1 pair from the high-variability condition versus 25% in the other 2 variability conditions. An exemplar-based learning account might explain the results. High token frequency could potentially facilitate storage of individual sentence exemplars, although the goal of the task was to train the participants to learn nonadjacent dependencies. In effect, language-impaired participants performed better in the condition where token frequency was highest, suggesting that at least some of these individuals were just memorizing individual strings (i.e., rote learning). As a result, the relatively higher token frequency in the low-variability condition became the most facilitative learning condition for these individuals. Again, caution is needed in interpreting these pattern differences because of limited statistical power, but the observed patterns reveal that some of the language-impaired adolescents might have learned the materials in a way different from the typically developing adolescents, but consistent with literature reviewed above suggesting exemplar-based language learning in SLI.

The Procedural Deficit Hypothesis of SLI

Recently, Ullman and Pierpont [2005] put forward a hypothesis of SLI as a procedural learning deficit. This procedural deficit hypothesis (PDH) is an extension of Ullman’s declarative-procedural model of language [Ullman, 2001] which postulates a distinct procedural learning system composed of interconnected brain structures, the most important of which appear to be the frontal/basal ganglia circuits [for a review, see Ullman & Pierpont, 2005]. According to Ullman [2001], the procedural memory system is important for learning of syntax and phonology, and contrasts with a declarative learning system that is involved in acquisition of vocabulary and more general semantic knowledge.

Historically, the concept of procedural learning was heavily influenced by early studies of neuropsychological patients, especially those with amnesia. Early studies focused on motor skill acquisition, with a striking demonstration that the famous patient H.M., who was densely amnesic after bilateral hippocampectomy, could master new motor learning, even though he could not remember doing the task [Corkin, 1968]. In a review, Squire [2004] noted that such early observations of preserved motor learning in amnesia were soon supplemented with accounts of spared perceptual learning, and visual and orthographic fragment completion (identifying pictures or words on the basis of partial cues). Of particular interest here were findings that people with amnesia could perform well on tasks involving extraction of prototypes and categories from probabilistic information, and on artificial grammar learning. Does this mean, then, that a single procedural system is involved in all aspects of implicit learning (fig. 1)? Squire [2004] argued against this idea, and proposed instead that there are multiple nondeclarative learning systems. There is evidence, for instance, that learning an artificial language might not draw on the same brain structures as skill learning, as patients with basal ganglia dysfunction can accomplish artificial grammar learning [Reber & Squire, 1999; Witt, Nühsman, & Deuschl, 2002]. With regard to SLI, Ullman and Pierpont [2005] argued that the deficit was in a procedural memory system that controlled learning and control of motor skills such as typing or riding a bicycle, and they drew attention to the motor deficits commonly found in SLI as evidence for their account. Their view of procedural learning is broader than that of Squire, [2004], who identified ‘procedural learning’ with mastery of motor skills and habits, and they postulated a deficit in extracting patterns from verbal material, which also extends to include all types of sequence learning, ‘serial or abstract, or sensorimotor or cognitive’. In terms of figure 1, then, the proposed deficit in the PDH encompasses the region encircled by the dotted line, possibly also extending to cover additional aspects of statistical learning.

The PDH generates 2 types of prediction. First, it predicts that children with SLI will show impairments in implicit sequence learning tasks, whether verbal or nonverbal. Second, the PDH implies that the nature of the language learning problem in SLI has to do with implicit learning of underlying structure from statistical features of the input, rather than rule extraction.

Empirical findings in line with the first prediction came from studies using nonverbal materials to test procedural learning in SLI. Using a serial reaction time task, Tomblin, Mainela-Arnold, and Zhang [2007] tested a group of adolescents with SLI who implicitly learned a 10-item repeating sequence by tracking a stimulus that moved among 4 spatial locations on a computer screen. Participants pressed 1 of 4 buttons on a response panel that matched the location where the visual stimulus appeared. Implicit knowledge of the sequence was revealed by faster reaction times when subsequently tested on the repeating sequence versus a random sequence. The adolescents with SLI showed a slower learning rate than typically developing controls. Poor procedural learning was observed by Lum, Gelec, and Conti-Ramsden [2009] who used a similar serial reaction time task and found poor procedural learning of sequence in a group of younger children with SLI¹ compared with age-matched typically developing children. The same children were unimpaired relative to typically developing children on a declarative learning task provided nonverbal materials were used. The relationship between grammatical abilities and procedural learning was demonstrated in a further analysis by Tomblin et al. [2007] who reclassified their participants into groups based on their grammar and vocabulary abilities and found a slower learning rate in the SLI group only when the language impairment was defined in terms of grammatical impairments, suggesting that the same learning mechanisms underpinning motor learning of sequential patterns might also be involved in learning grammar.

Turning to consider the nature of the mechanism of implicit language learning, Perruchet and Pacton [2006] recently compared the ways in which the terms statistical and implicit learning had been used. In their article, they did not consider neuropsychological bases of different implicit memory systems. Instead, implicit learning was used to refer to the general ability to learn sequential patterns when there is no conscious intent to extract this sequential information. The authors argued that although implicit learning and statistical learning have much in common, researchers in these 2 fields tend to emphasize different learning processes. In the implicit learning literature, the focus has been on the way in which grammaticality judgments in artificial language learning tasks rely on knowledge of fragments of strings or chunks [Knowlton & Squire, 1996], and the same process could explain reductions in response time in serial reaction time tasks [Jiménez, 2008]. In contrast, in most artificial grammar research, the emphasis has been on how humans learn general statistical information. There is, however, a confusion between predictability of stochastic patterns and chunk strength, and grammatical test strings will typically have greater chunk strength than ungrammatical strings.

Recent imaging studies have suggested that both chunk strength and statistical properties of input are used in implicit learning. Liberman et al. [2004] used eventrelated functional magnetic resonance imaging to identify neural regions involved in artificial grammar learning. They controlled for the chunk strength effect, by ensuring that the average chunk strength of grammatical items was equivalent to the average chunk strength of ungrammatical items. They found that activation in the right caudate was associated with pattern learning, whereas medial temporal lobe activations were associated with chunk strength. These findings suggest that both chunk formation and statistical learning of patterns are involved in artificial grammar learning. Interestingly, imaging studies have also provided converging evidence that both striatum [Poldrack et al., 2001; Rauch et al., 1995] and medial temporal lobe [Curran, 1997; Schendan et al., 2003] are recruited during serial reaction time tasks.

At this point, we may ask whether the PDH is the same as a statistical learning account of SLI. Although there is considerable overlap in these viewpoints, they are not identical. The PDH adopts a more polarized view whereby different aspects of language learning are mediated by declarative and procedural systems, which have separate neurobiological bases. Thus the declarative system is seen as handling not only learning of vocabulary and semantic information, but also other arbitrary linguistic information that is not predictable from hypothesized rules, including storage of irregular past forms [Pinker & Ullman, 2002]. In contrast, research in the field of statistical learning has converged on a conclusion that statistical learning mechanisms can achieve grammatical acquisition as well as other aspects of language, such as learning arbitrary associations between sounds and referent [Smith & Yu, 2008] and generalizing names for solid objects by shape in learning vocabulary [Samuelson, 2002]. Another example that demonstrates the differences between the 2 accounts is English past tense morphology. Ullman and Pierpoint [2005] draw a clear line between memory systems underpinning acquisition of irregular past tense (declarative learning) and regular past tense (procedural learning), citing in support disproportionately poorer performance by children with SLI in regular past tense compared with irregular past tense [e.g., van der Lely & Ullman, 2001]. However, a recent study has queried this conclusion, showing that when input frequency is controlled, children with SLI make as many errors in marking past tense for regular verbs as for irregular verbs [Serratrice, Conti-Ramsden, & Joseph, 2003]. In addition, there is evidence that stochastic regularities in the input are sufficient to achieve acquisition of both regular and irregular past tense [Albright & Hayes, 2003].

Impact of Limitations in Perception or Short-Term Memory on Statistical Learning

Demonstration of poor statistical learning does not necessarily indicate that this is the core deficit in SLI. It could be that extraction of statistical structure from language input is poor because the incoming information is not adequately perceived. Suppose, for instance, that a child had difficulty in discriminating between all the phonemes in the native language. This could have 2 adverse consequences for statistical language learning. First, some key regularities in the input may be missed: for instance, it has been noted that in English, past tense and plural endings are brief and may be hard to perceive [Leonard, 1989]. This could make it harder to extract regularities involving these endings. Second, if minimal pairs of words are perceived but not adequately distinguished, then the number of perceived types for a given number of tokens, X, in a structure such as A-X-B will be reduced. As noted above, this would make it harder to extract the A-B regularity. Although we cannot rule out this kind of mechanism, it does not seem a sufficient explanation for poor language learning in SLI. First, the syntactic problems of children with SLI are markedly worse than of children with mild-to-moderate hearing loss [Norbury et al., 2001]. Second, poor comprehension of grammatical contrasts by children with SLI was found in a training study by Bishop and colleagues [2006], even in optimal listening conditions with simple, pictured vocabulary and short sentences. Furthermore, acoustic modification of the spoken sentences to lengthen and amplify brief and nonsalient portions of the speech signal had no beneficial effect. Finally, this account could not explain deficits in statistical learning in nonverbal motor tasks [Tomblin et al., 2007].

Limitations of short-term memory could, however, potentially be implicated in weak statistical learning. Abstract syntactic patterns are built over individual exemplars that are stored in the first place. In order to detect a pattern in input such as A-X-B, A-Y-B, A-Z-B, one must be able to retain at least some 3-element sequences in memory. If the memory span is only 2 items long, the pattern will not be detected. Suppose a child with SLI retains only 50% of 3-element utterances, whereas a typically developing child retains 90% of the same material. This would mean that children with SLI require more exposure in order to memorize a sufficient amount of materials before more abstract patterns could emerge, and so language will be more dependent on memorizing chunks of words. Given that deficits in working memory in SLI are well attested in the literature [for a review, see Coady and Evans, 2008], it is possible that the seeming overreliance on rote learning is a result of limited memory capacity. This suggests that rather than a deficit in statistical generalizations, the observed poor performance in statistical learning could reflect a fundamental difficulty in reaching a ‘critical mass’ [Marchman & Bates, 1994]. This interpretation makes testable predictions. Consider the nonadjacent dependency task, for instance. If children with SLI have adequate statistical learning abilities, but their learning is hampered by deficient working memory, they should be slower than other children at reaching an adequate level of performance in grammatical judgment of previously heard sentences, but once that level is reached, they should show normal performance with novel sentences. On the other hand, if working memory does not explain poor statistical learning in SLI, then children with SLI should reach an adequate level of grammaticality judgment of sentences they heard during training (exemplar-based), but accurate judgment of novel sentences (generalization) would still be challenging.

Corpus analyses provide another possible way to compare these accounts. As high token frequency would facilitate exemplar learning, type frequency is critical for generalization. It is interesting to explore adult language input and test whether children with SLI at a young age produce utterances of relatively high token frequency to a similar level as typically developing children and, more importantly, whether they do so for a protracted period. On the other hand, a syntactic construction that is high on type frequency and low on token frequency should pose more challenges for children with SLI than a syntactic construction of a similar type frequency but a relatively higher token frequency.