Both domain-specific mechanisms, such as phonological processing (Bailey & Snowling, 2002), and domain-general mechanisms, such as working memory (Gathercole & Baddeley, 1993), have been implicated in language development, and in word learning in particular. Learning an unfamiliar word taps multiple cognitive processes, including rapid generation of a lexical representation, retention of the serial order of constituent phonemes, and creation of long-term phonology-to-semantics connections (Gupta & MacWhinney, 1997). Of particular interest in the present study was the acquisition of novel phonological forms and sequences, a task considered to place high demands on both immediate memory processes and phonological processing. In the present study, we explored the relative influences of domain-general working memory and domain-specific phonological abilities on word learning by contrasting immediate recall and learning across visual and auditory modalities, as well as with familiar and unfamiliar phonological forms.

We investigated domain-specific and domain-general constraints on word learning here by comparing the performance of children with developmental impairments in either language learning or working memory. Specific language impairment (SLI) refers to an unexpected delay in the onset or development of language, despite otherwise typical development and opportunities. Domain-specific deficits affecting all aspects of language learning are well documented for SLI groups, including impairments in phonological processing (Vandewalle, Boets, Ghesquière, & Zink, 2012) and vocabulary acquisition (Riches, Tomasello, & Conti-Ramsden, 2005). Although children with SLI perform poorly on working memory measures involving both the storage and manipulation of verbal materials (Archibald & Gathercole, 2006a), when working memory deficits are assessed using visuospatial measures, deficits are not consistently reported among SLI groups (Archibald & Gathercole, 2006b). In contrast, children with a specific working memory impairment (SWMI) perform poorly on both verbal and visuospatial working memory tasks but achieve age-appropriate scores on standardized language tests (Archibald & Joanisse, 2009). Our SWMI group was considered to have a domain-general working memory deficit because their impairments were established for working memory tasks involving the storage and processing of verbal and visuospatial materials, but not for short-term memory tasks requiring storage only.

Both domain-specific and domain-general processes have been described as part of Baddeley and Hitch’s (1974) model of working memory, or the ability to briefly store and process information. According to this model, working memory includes two domain-specific slave systems that handle the temporary storage of phonological and visuospatial materials (the phonological loop and the visuospatial sketchpad, respectively). The third component of working memory is the central executive, which is a controlled attentional resource responsible for high-level processing activities and for the coordination of activities within working memory. Suggestions that the central executive is a domain-general resource are supported by findings of a common processing efficiency factor underlying both verbal and visuospatial working memory tasks (Chein, Moore, & Conway, 2011; Kane et al., 2004). The children identified with SWMI in our previous study (Archibald & Joanisse, 2009) were considered to have a deficit in the domain-general central executive component of working memory because their impairments were established for working memory tasks involving the storage and processing of verbal and visuospatial materials, but not for short-term memory tasks requiring storage only.

Short-term and working memory tasks are distinguished by whether they involve the brief retention of information in short-term memory only (e.g., serial recall of words) or involve additional processing involving the central executive (e.g., reversing list order prior to recall). As well, serial recall tasks involving long lists that exceed an individual’s short-term memory capacity share common variance with other complex working memory tasks (Unsworth & Engle, 2006). It has been suggested that such supraspan tasks additionally tap domain-general resources in working memory to retrieve items displaced from short-term memory (Unsworth & Engle, 2006). Both simple span measures of phonological short-term memory (Gupta & MacWhinney, 1997) and complex verbal working memory tasks (Daneman & Carpenter, 1983) have been linked to language abilities. Visuospatial short-term memory tasks, however, have not been found to be related to language processing (Adams, Bourke, & Willis, 1999; but see Phillips, Jarrold, Baddeley, Grant, & Karmiloff-Smith, 2004). As a result, in the present study we focused on the domain-specific processes of phonological short-term memory and the domain-general resources of working memory.

Domain-specific phonological processes involving the retention, coding, and manipulation of phonological information have been found to be highly related to the long-term retention of phonological forms in word learning (Gupta & MacWhinney, 1997; Leclercq & Majerus, 2010; Majerus et al., 2009; Majerus, Poncelet, Greffe, & Van der Linden, 2006). The phonological loop, as described by Baddeley and Hitch (1974) and subsequently developed by others (Burgess & Hitch, 1998; Page & Norris, 1998), incorporates both the storage of phonological material in short-term memory and its recoding from visual stimuli. Phonological short-term memory is traditionally tested in immediate serial recall tasks, with responses forming a classic serial position curve in which recall starts very accurately, decreases throughout the list, and then improves toward the end of the list (e.g., Murdock, 1962; Waugh & Norman, 1965). Strong links have been reported between measures of phonological short-term memory and children’s learning of novel phonological forms and vocabulary development (Baddeley, Gathercole, & Papagno, 1998; Jarrold, Baddeley, Hewes, Leeke, & Phillips, 2004; Munson, Kurtz, & Windsor, 2005). Conversely, lexical knowledge also supports retention in phonological short-term memory, in that recall is better for familiar words (Gathercole, Frankish, Pickering, & Peaker, 1999) or nonwords sharing features of real words (Gathercole, 1995). Nevertheless, poor recall on phonological short-term memory tasks alone has not been found to be sufficient to account for long-term deficits in language learning (Briscoe, Bishop, & Norbury, 2001; Gathercole, Tiffany, Briscoe, Thorn, & the ALSPAC Team, 2005). This pattern of findings has led to the suggestion that phonological short-term memory is important for the long-term retention of novel phonological sequences, especially during the early stages of vocabulary learning, but that learning at later stages is supported by the amassed lexical store and linguistic knowledge (Gathercole, 2006).

The present study focused on one aspect of phonological processing that has been suggested to be part of the phonological loop: the ability to recode phonological forms from visual material. Comparing the recall of equivalent visually and auditorily presented lists allowed us to focus on the efficiency of the recoding process. We hypothesized that the extra demands of phonological recoding might impair the performance of our SLI group with domain-specific language-learning deficits, but not of our SWMI group with domain-general central executive impairments. It should be noted that the role of group differences in long-term lexical knowledge in the present study was minimized by using either nonwords or highly familiar and frequent items that were known and easily named by all participants. In this way, our study focused on word and sequence learning that was comparably novel for all of the children.

Other processes in addition to phonological processing undoubtedly influence new word-form learning. While phonological short-term memory acts to briefly retain rapidly encoded phonological representations, other questions are involved with how long-term, ordered phonological representations are formed as stable representations within the lexical system. Presumably, with repeated exposures connections between novel phonological sequences become sufficiently strong to resist decay over time (Gupta, 2003). One possibility is that domain-general cognitive abilities related to attentional and executive processes mediate the creation of long-term lexical representations (Bayliss, Jarrold, Baddeley, Gunn, & Leigh, 2005; Jones, Farrand, Stuart, & Morris, 1995; Mosse & Jarrold, 2008) and facilitate connections within a semantic network (Mosse & Jarrold, 2008). It has been suggested that working memory may play such a role. This notion was investigated in the present study by examining the links between working memory and word and sequence learning.

Experimental demonstration of long-term sequence memory comes from the Hebb repetition effect, so termed after Hebb’s (1961) original report that immediate serial recall in a block of trials improves to a greater extent on a list repeated about every three trials, as compared to nonrepeated filler lists. Differences in performance between repeated “Hebb” trials and the nonrepeated “filler” trials reflect a dichotomy between short-term sequence memory for immediate serial recall and (implicit) long-term sequence memory for the Hebb trials. Working memory is tapped in the Hebb paradigm because it is usually presented as a supraspan task, in which the length of the lists presented is just beyond that which an individual can consistently recall immediately from short-term memory (Gagnon, Bédard, & Turcotte, 2005; Turcotte, Gagnon, & Poirier, 2005). According to the working memory model (Baddeley & Hitch, 1974), domain-general central executive resources are recruited when demands exceed the capacity of one of the domain-specific short-term memory stores.

The Hebb effect has been replicated in many studies involving adults (e.g., Couture & Tremblay, 2006; Cumming, Page, & Norris, 2003), but only more recently in typically developing children (Mosse & Jarrold, 2008) and children with impairments (Majerus et al., 2008; Mosse & Jarrold, 2010). In Mosse and Jarrold’s (2008) study, a group of 5- to 7-year-old typically developing children completed three blocks of trials in which Hebb lists were presented alternately with eight filler lists. In order to demonstrate a divergence in performance across successive trials on Hebb as compared to filler lists, their analyses compared first- and second-half block scores. Hebb effects were reflected by the maintenance of Hebb trial scores in the context of a reliable decline in filler-trial performance. Mosse and Jarrold (2008) suggested that the filler-list decrement was likely due to fatigue effects or to the buildup of proactive interference in their young participants, effects that were counteracted by Hebbian learning on the repeated lists. No group differences in Hebbian learning, when compared to typically developing matched controls, have been found for SLI (Majerus et al., 2008) or Down syndrome (Mosse & Jarrold, 2010) groups. These findings have been interpreted as suggesting that basic serial-order detection and learning mechanisms are preserved in both SLI and Down syndrome, despite deficits in phonological short-term memory in these groups. Consistent with these findings, then, a more domain-general mechanism such as working memory may facilitate long-term sequence learning (Bayliss et al., 2005; Jones et al., 1995; Mosse & Jarrold, 2008). We examined the importance of domain-general working memory abilities to long-term sequence learning in the present study by comparing Hebb repetition performance among children with SWMI or SLI and typically developing children. To our knowledge, Hebb repetition effects have not been examined previously in children with SWMI. Indeed, no prior studies have directly investigated working memory influences in Hebbian learning, either by examining individual differences in typical learners or by comparing performance in individuals with a specific impairment in domain-general aspects of working memory, such as SWMI.

Evidence has also suggested that the Hebb effect operates across modalities (Page, Cumming, Norris, Hitch, & McNeil, 2006). In adults, Page et al. investigated cross-modal Hebb effects from lists presented in either the auditory or visual modality. Items presented in the visual modality included letters and pictures and were easily labeled verbally. A reliable transfer of Hebbian learning was observed from repeated lists that had previously been presented in either the auditory or visual modality to equivalent repeated lists presented in the opposite modality. In the present study, we examined Hebb effects across modalities. This manipulation addressed our hypothesis that the extra demands of phonological recoding in long-term sequence learning might impair the performance of our SLI group with domain-specific language-learning deficits, but not of our SWMI group with domain-general central executive impairments.

We further examined the roles of domain-specific and domain-general mechanisms in a paired-associate learning task involving the explicit pairing of a phonological form and an image. The paired-associate learning paradigm differs from the Hebb effect because it involves explicit learning, a conscious intention to learn (Ellis, 1994; Hulstijn, 2001). Nevertheless, it has been suggested that the sequence-learning mechanism tapped experimentally in the Hebb repetition paradigm supports word learning generally (Page & Norris, 2009; Szmalec, Duyck, Vandierendonck, Mata, & Page, 2009). Mosse and Jarrold (2008) employed an adapted version of the paired-associate task involving word–picture and nonword–picture pairs in their group of typically developing 5- to 7-year-olds. Learning of the nonword–picture but not the word–picture pairings was reliably correlated with their Domain-Specific Phonological Short-Term Memory factor, whereas implicit learning measured in a Hebbian-learning task was correlated with both word and nonword associations. These findings suggest that implicit learning supports explicit learning generally, with the learning of words and nonwords placing similar demands on domain-general memory mechanisms. Phonological processing, on the other hand, is tapped o a greater extent in the associative learning of novel than of familiar phonological forms, because retention of the familiar forms is supported by long-term lexical knowledge (Gathercole, 1995; Munson, 2001). In the present study, we hypothesized that children with SWMI might show decreased associative learning for both words and nonwords due to their domain-general working memory deficits. While we expected that children with SLI might show reduced paired-associate learning, as has been reported for other SLI groups (Lum, Gelgic, & Conti-Ramsden, 2010), we hypothesized that the SLI group would be disproportionately impaired in the nonword condition.

In the present study, school children with typical development or a developmental language and/or working memory impairment completed a paired-associate learning task involving familiar and novel phonological forms, in addition to a Hebbian learning paradigm involving serial recall of visual or auditory lists in which uniquely ordered filler lists were alternated with a fixed-order Hebb list. The visual and auditory lists in the Hebb paradigm were equivalent (i.e., the same items) in order to allow us to focus on the phonological processing required for cross-modality recoding and carryover from the visual to the auditory modality. As well, the lists in the Hebb paradigm were presented at a supraspan level equivalent to one item more than each respective participant’s digit recall span, thereby equating the phonological short-term memory demands across participants and tapping working memory.

Overall, we anticipated that working memory would constrain learning generally, while language abilities would impact phonological learning. More specifically, one aim of the study was to examine the domain-specific phonological processing constraints on word learning across paradigms by comparing the performance of SLI and typically developing groups. We expected to see evidence of Hebbian learning generally in the SLI group. However, we hypothesized that the children with SLI would be less skilled at phonological recoding, and thus would exhibit reduced carryover learning from the visual to the auditory modality. Additionally, a disproportionate SLI deficit in paired-associate learning was anticipated that would involve novel as compared to familiar phonological forms.

A second goal was to evaluate domain-general working memory influences on word learning across paradigms by comparing the performance of SWMI and typically developing groups. We hypothesized that children with SWMI would show reduced Hebbian learning and paired-associate learning generally that would affect performance in the visual and auditory modalities similarly, as well as affect the learning of both familiar and novel phonological forms.

Method

Participants

All of the children had participated in our previous study investigating language and working memory impairments in school age children (Archibald & Joanisse, 2009). Invitations were extended to 74 children, including all of those identified in our previous study (except one who could not be located) as having combined language and working memory impairments or specific impairments in either language or working memory. The recruitment rates for children in the impaired groups was 69 %, and for the typically developing group, 85 %. A total of 58 children (29 boys, 29 girls, age range 6.3–10.2 years) participated in the present study; of these, 12 met the criteria described below for a language impairment (LI) only, eight for a working memory impairment (WMI) only, and 11 for both LI and WMI. It was necessary to include children with comorbid deficits in language and working memory in order to recruit sufficient sample sizes in our LI and WMI groups. Overall, then, 23 children had an LI (± WMI), and 19 had a WMI (± LI). A total of 27 had no LI or WMI (typical development; TD).

All of the children had previously completed a battery of standardized tests 4–5 months prior to the present study (described in detail in Archibald & Joanisse, 2009). Briefly, the Test of Nonverbal Intelligence 3 (TONI-3; Brown, Sherbenou & Johnsen, 1997) was administered as a measure of general nonverbal cognitive ability. In addition, the four core subtests of the Clinical Evaluation of Language Fundamentals 4 (CELF-4; Semel, Wiig, & Secord, 2003) were completed as the reference standard for language skills. The core subtests consisted of Concepts and Following Directions, Recalling Sentences, Formulating Sentences, and, depending on the age of the child, Word Knowledge (under nine years) or Word Classes (nine or older). Participants were considered to have a language impairment (LI) if their composite language score on the core subtests of the CELF-4 was more than 1 SD below the mean (<86). A cutoff of 1 SD below the standardized mean had been used in many previous studies (e.g., Bishop & Norbury, 2002; Rice, Wexler, Marquis, & Hershberger, 2000) and corresponds to an effect size of 1.0, conventionally considered to be large in magnitude (Cohen, 1988).

The Automated Working Memory Assessment (AWMA; Alloway, 2007) provided a test of working memory. The AWMA includes 12 subtests, three of which target each of phonological short-term memory (recall lists of digits, words, or nonwords), visuospatial short-term memory (recall locations of dots, blocks, or paths through a maze), verbal working memory (recall tallies, digits, or final words after counting, reordering, or processing a sentence, respectively), and visuospatial working memory (recall location or orientation after identifying a different shape or mentally rotating an image, respectively). Participants were considered to have a WMI if their average standard score across both the visuospatial and verbal working memory composite scores was more than 1 SD below the mean. Table 1 presents descriptive statistics for the participants, grouped according to the presence or absence of LI or WMI. The impaired groups scored lower than the typically developing groups on all measures. Group differences in language or working memory that were not of primary interest in the individual analyses (i.e., working memory, in an analysis focused on language) were taken into account in the present study by adding the respective factor as a covariate in the statistical analysis (i.e., adding working memory as a covariate in an analysis focused on language).

Table 1 Descriptive statistics for standardized tests (M = 100, SD = 15) of language, nonverbal intelligence, and short-term memory (STM) and working memory (WM) for all groups
Full size table

Procedure

All participants completed three individual sessions of 30–40 min approximately one week apart, in a quiet room in their school or at our research laboratories; these sessions included the tasks described in the present study and others not reported here. The digit span task was completed in the first session, the paired-associate learning task in the second session, and the Hebb visual–auditory learning task in the third session.

The digit span task was administered in order to measure each child’s immediate serial recall span. In the digit span task, the child was required to repeat sequences of digits in correct serial order, presented at a rate of one digit per second. The digit lists were random constructions without replacement from the digits 1 through 9, recorded in a female voice, and presented via E-Prime (Version 1.1; Psychology Software Tools, Sharpsburg, PA). All children began with two trials at a list length of two digits and then continued, completing three trials of three-digit lists, and so on, to a maximum list length of seven. Testing was discontinued when the child failed to accurately repeat three trials at any one list length, and this point was considered the child’s “span + 1” list length or supraspan level. All of the Hebb tasks were completed at each respective participant’s supraspan level in order to equate phonological short-term memory demands across all participants. In this way, group differences in implicit learning could not be attributed to a difference in the short-term memory load imposed by the task. Table 1 includes descriptive statistics for digit span performance. Although both impaired groups completed fewer correct trials in the task than did the typically developing group, there were no differences in mean supraspan length for children with or without WMI (p > 0.05) or children with or without LI (p > 0.05).

In the Hebb visual–auditory learning task, each child completed three blocks at their supraspan level as determined by the digit span task. A block consisted of 16 visual trials, followed immediately by 12 auditory trials. Three sets of word lists were generated at each supraspan length required (four-, five-, six-, and seven-item lengths). For each block, a single group of words of the required set size (i.e., a single group of four, five, six, or seven words) was chosen randomly from a closed set of common one-syllable imageable nouns (bath, bus, cake, clock, doll, fish, girl, hand, horse, leaf, man, spoon, train), with the constraints that no duplicates occurred within a set, and only the minimum number of duplicates across the three groups required for each length. For each word group selected for a set, 15 uniquely ordered lists were generated by sampling the words without replacement. Of these lists, one was selected to be the repeated list (referred to as the Hebb list) that was presented as a visual (picture) list on the first eight even-numbered trials and as an auditory list on the final six even-numbered trials. Of the remaining lists (referred to as filler lists), eight were selected to be nonrepeated visual lists, which were presented on the first eight odd-numbered trials. The remaining six filler lists were presented as auditory lists on the subsequent six odd-numbered trials. Thus, we produced three sets of lists (blocks) at each of four supraspan lengths, each set comprising 16 visual lists and 12 auditory lists, with the Hebb list repeated 14 times on even-numbered trials (eight visual, six auditory), and 14 filler lists (eight visual, six auditory) each appearing once (odd trials). In this study, Hebb lists were presented on alternate trials rather than the more traditional every third trial because alternate-trial presentations have been more effective in demonstrating Hebbian-learning effects in young children (e.g., Mosse & Jarrold, 2008, 2010). Only six Hebb and six filler lists were completed in the auditory modality (rather than eight each, as in the visual modality) because pilot testing had revealed that children were better able to maintain attention with fewer trials. In order to ensure that Hebbian learning occurred in the visual trials, we presented the same number of Hebb trials as had been employed by Mosse and Jarrold (2008). Nevertheless, we anticipated a large effect size for the auditory advantage, due to carryover of learning from the visual modality and the obligatory access of acoustic stimuli to short-term memory, making it possible to detect the auditory advantage with fewer trials.

For the visual trials, a single, stock color photograph depicting each word in the sequence was presented in the center of a white background for 1 s. After the list, all of the pictures appeared in a fixed circular display, and the child pointed to the pictures in the correct serial order. Although children were not instructed to reply verbally, full credit was given for accurate verbal responses. The corresponding auditory trials followed the same format. The spoken words were recordings of a female voice presented at a rate of one per second, after which the child attempted to recall each word in sequence verbally. The children were given short breaks between the three blocks. The children were not told that any of the sequences repeated, nor were they asked if they noted any repeating sequences. The percentages of items recalled in the correct serial position were calculated.

The paired-associate learning task was based on the task reported by Mosse and Jarrold (2008). Two sets of four novel creatures were chosen from the “space aliens” set of Gupta et al. (2004). In the word-learning trials, each of four of the creatures was paired with a familiar name (Michael, Simon, Thomas, and Peter), and learning of these associations was measured. On any trial, the four aliens appeared one at a time in the center of the computer screen in a fixed random sequence with no sequences repeated across trials, and a female voice recording of the name of each alien was played by the computer. All four creatures then appeared on the screen at once in a fixed 2 × 2 configuration, and the child was asked to name each one, with the number of correct responses being recorded. A total of ten trials were conducted, unless all four aliens were correctly named in two consecutive trials, whereupon testing ceased. Individuals who did not complete all ten trials due to maximum scoring on two consecutive trials were credited with full recall on the remaining, untested trials. The nonword-learning trials were conducted after the word-learning trials and in exactly the same manner as the word-learning trials, except that nonwords rather than familiar names were employed. The nonword names were created by Mosse and Jarrold (2008) by rearranging the phonemes of the four familiar names used in the word-learning trials, to produce Meton, Sommel, Tiker, and Pimus.

Data analysis

An arcsine square-root transformation was completed on all percent scores in order to transform the fixed-limit distribution of percentages to a normal distribution appropriate for statistical analyses (Studebaker, 1985), although, for clarity, all descriptive statistics presented in the tables and figures represent percent correct scores. Developmental patterns for the typically developing group were assessed using ANOVAs for serial recall in the Hebb paradigm, implicit Hebbian learning, and paired-associate explicit learning. The impact of an SLI was examined in an ANCOVA comparing groups with or without LI while controlling for individual differences in working memory and nonverbal intelligence. Corresponding ANCOVAs with language and nonverbal intelligence scores entered as covariates examined the impact of an SWMI. The nonverbal intelligence measure was entered as a covariate in these analyses in order to account for cognitive differences present in a large cross-sectional developmental sample. In cases of a violation in the assumption of equality of covariance matrices, we instead completed a stepwise regression.

In order to evaluate implicit learning in the Hebb tasks, we employed the procedure adopted by Mosse and Jarrold (2008) of comparing performance on the first and second halves of each list type. While a main effect of list type in favor of the Hebb sequence might provide some evidence of implicit learning, only the demonstration of improvements in performance over repeated Hebb list presentations would clearly reflect implicit learning. Thus, an interaction between list and half due to higher scores on the Hebb list over the second half of trials would be needed to provide clear evidence of implicit learning.

Results

Hebbian learning across trials

Typical development (TD)

For the initial analysis, we examined implicit sequence learning in our TD group. Table 2 summarizes descriptive statistics for items correct in the first or the second half of the visual versus auditory Hebb and filler lists, averaged across the three Hebbian-learning blocks. In the 2 (modality: visual / auditory) × 2 (list type: filler / Hebb) × 2 (half: first/second) ANOVA completed on the transformed data, all effects but one were significant: list, F(1, 26) = 39.43, p < 0.001, \( \eta_p^{2} = 0.60 \); half, F (1, 26) = 9.28, p = 0.005, \( \eta_p^{2} = 0.26 \); Modality × List, F (1, 26) = 7.73, p = 0.01, \( \eta_p^{2} = 0.23 \); Modality × Half, F (1, 26) = 9.69, p = 0.004, \( \eta_p^{2} = 0.27 \); List × Half, F(1, 526) = 8.45, p = 0.007, \( \eta_p^{2} = 0.24 \); Modality × List × Half, F (1, 57) = 9.155, p = 0.004, \( \eta_p^{2} = 0.14 \). Figure 1 illustrates the three-way interaction of modality, list, and half. If we first consider the visual trials, we observed the maintenance of Hebb trial scores and a decrement on filler trials exactly matching that reported by Mosse and Jarrold (2008). On the auditory trials, Hebb scores were significantly higher than those for the visual trials, reflecting carryover learning from the visual trials. Performance on the first half of filler scores was significantly lower in the auditory than in the visual trials (p = 0.021), a difference that did not occur in the second half of trials (p > 0.05). As a result, the pattern of a decrement on the second half of filler trials was not observed for the auditory trials. The main effect of modality was the only nonsignificant effect in this analysis, F (1, 26) = 0.96, p > 0.05.

Table 2 Mean percentages of items correct (with standard errors) for each learning task and participant group
Full size table
Fig. 1
figure 1

Mean percentages of items correctly recalled (with standard errors) across visual and auditory trials for Hebb and filler lists, by list halves

Full size image

Specific language impairment

We next turn to the effect of SLI on Hebbian learning. Grouped data for the Hebbian-learning task are shown in Table 2. An ANCOVA was completed for the effects of group (presence/absence of SLI), modality (visual/auditory), list type (filler/Hebb), and half (first/second), with working memory composite and nonverbal intelligence scores entered as covariates. The results revealed two significant interactions with group: the interaction between modality and group, F (1, 54) = 4.444, p = 0.040, \( \eta_p^{2} = 0.08 \), and that between half and group, F (1, 54) = 5.899, p = 0.019, \( \eta_p^{2} = 0.10 \). All remaining effects and interactions involving group were not significant, F < 1.4, p > 0.05, as were all effects involving nonverbal intelligence (F < 2.1, p > 0.05, all cases) and Box’s test of the equality of covariance matrices, M = 50.54, p > 0.05. Figure 2 displays the interaction between modality and group, such that the children with SLI performed significantly more poorly than did the group without SLI in the auditory condition. Importantly, this interaction was not differentiated by list types, indicating a general auditory retention difficulty rather than a specific deficit in carryover learning on the Hebb lists. Figure 3 shows the interaction between group and list half, reflecting better retention in the first half of lists for the children without SLI. Once again, this pattern was not differentiated by list type, so it cannot be accounted for by group differences in Hebbian learning.

Fig. 2
figure 2

Estimated mean (with standard error) percentages of items correct (adjusted for composite working memory and nonverbal intelligence) for participants with or without specific language impairment (SLI) in the visual versus auditory modalities (collapsed across list types and halves), depicting better performance in the auditory than in the visual Hebb lists for the typical-language but not for the SLI group.

Full size image
Fig. 3
figure 3

Estimated mean (with standard error) percentages of items correct (adjusted for composite working memory and nonverbal intelligence) for participants with or without specific language impairment (SLI) in the first versus the second half of lists (collapsed across modalities and list types), depicting better performance in the first half of lists for the typical-language than for the SLI group

Full size image

Specific working memory impairment

We next examined the effect of SWMI on Hebbian learning. A corresponding ANCOVA performed on the Hebbian-learning trials with group (presence/absence SWMI), modality (visual/auditory), list type (filler/Hebb), and half (first/second) as variables, and the composite language and nonverbal intelligence scores entered as covariates, examined the impact of SWMI on implicit sequence learning. We found no significant main effects or interactions with group (F < 2.35, p > 0.05, all cases) or nonverbal intelligence (F < 1.9, p > 0.05, all cases). The Box’s test of the equality of covariance matrices was not significant, M = 43.07, p > 0.05.

To summarize, Hebbian learning in the typically developing group was clearly apparent in the visual trials, with maintenance of Hebb performance and a decrement on filler trials. Carryover learning from the visual to the auditory Hebb lists was reflected in the significantly higher performance on the auditory Hebb lists than on the visual Hebb lists or on the filler auditory lists. Performance was significantly lower on the first half of the auditory than of the visual filler trials, with no subsequent decrement on the auditory filler trials. Children with SLI performed significantly more poorly in the auditory condition and on the first half of trials across list types. The performance of children with SWMI was not differentiated from that of typically developing children in this implicit sequence-learning task with phonological short-term memory demands equated. Nonverbal intelligence did not interact with learning in the Hebb visual–auditory paradigm.

Paired-associate learning

Typical development

We next examined performance on the word and nonword paired-associate learning tasks in the TD group (Table 3). As expected, scores were significantly higher for word than for nonword learning, t (26) = 6.16, p < 0.001.

Table 3 Correlations between paired-associate learning, implicit learning, second-half performance for filler lists, and standardized test scores
Full size table

Specific language impairment

A 2 (group: presence/absence of SLI) by 2 (name: word/nonword) ANCOVA with the composite working memory and nonverbal intelligence scores entered as covariates could not be completed on the paired-associate learning scores, due to a violation in the assumption of the equality of covariance matrices, Box’s M = 12.37, p = 0.008. We investigated language effects by completing stepwise linear regressions on the total items-correct scores separately for each name type, with the composite language, working memory, and nonverbal intelligence scores entered as covariates. The model for the word learning context was significant, F (1, 56) = 6.695, p = 0.012, accounted for 9 % of the variance, and retained only the working memory composite, whereas the significant model for the nonword learning context, F(1, 55) = 12.557, p < 0.001, accounted for 29 % of the variance and included both the language and working memory composites, which explained 7.9 % and 6.9 % of the unique variance, respectively. Nonverbal intelligence was not retained in either model. Taken together, these results indicate that poor language abilities were associated with reduced paired-associate learning of nonword–picture pairings specifically (Fig. 4).

Fig. 4
figure 4

Mean percentages of trials correct (with standard deviations) in the paired-associate learning task for words versus nonwords among the children with or without specific language impairment (SLI), reflecting disproportionate difficulty learning nonwords for the SLI group, F (1, 56) = 4.9, p = 0.031, \( \eta_p^{2} = 0.08 \)

Full size image

Specific working memory impairment

A significant Box’s test of the equality of covariance matrices was also observed in the corresponding ANCOVA investigating the impact of SWMI on paired-associate learning (M = 14.38, p = 0.013). The results of the regression analysis reported above indicate that poor working memory was associated with lower paired-associate learning scores, regardless of name type (Fig. 5).

Fig. 5
figure 5

Mean percentages of trials correct (with standard deviations) in the paired-associate learning task for words versus nonwords among the children with or without specific working memory impairment (SWMI), reflecting difficulty learning across modalities for the SWMI group, F (1, 57) = 15.1, p < 0.001, \( \eta_p^{2} = 0.21 \)

Full size image

Relationships between tasks and tests

In a final analysis, we examined the associations between the paired-associate learning tasks (word, nonword); implicit sequence learning, as reflected by performance on the second half of Hebb lists (visual, auditory); performance on the second half of filler lists, as an index of general verbal serial recall (visual, auditory); and the standardized test scores of nonverbal intelligence, language, and short-term and working memory. Zero-order correlations are presented in Table 3 for all participants.

One clear pattern emerging from the correlational analysis (Table 3) was that paired-associate learning of nonwords was highly correlated with all measures, including both visual and auditory Hebbian learning (r > 0.28, p < 0.05, all cases), whereas the paired-associate learning of words was associated with composite working memory (r = 0.33, p < 0.05) and visual Hebbian learning (r = 0.28, p < 0.05). The Hebb and filler tasks were highly correlated, as was expected given the similar serial recall demands of these tasks. Thus, explicit learning for words was reliably associated with domain-general working memory and Hebbian learning, while learning for nonwords was additionally related to phonological short-term memory and to language and general cognitive abilities.

Discussion

The primary aim of this study was to examine language- and memory-related constraints on word learning. We compared the performance of groups with and without developmental impairments in language and/or working memory on an explicit measure of children’s ability to learn new vocabulary and an implicit measure of sequence learning, the Hebb repetition effect. List length for serial recall in the Hebb task was set individually at one item more than the child’s digit recall span, thereby equating phonological short-term memory tasks across participants and tapping working memory. In addition, a Hebb manipulation involving the presentation of visual (picture) followed by equivalent auditory (spoken name) sequences allowed us to examine the impact of language or working memory on phonological recoding in the carryover of learning from the visual to the auditory modality. Implicit sequence learning, as evidenced by maintenance of Hebb trials and a filler list decrement, was observed in the visual modality for all participants, as was carryover from visual to auditory Hebb lists. Regardless of list type, children with an SLI performed more poorly in the auditory trials and on the first half of lists. The presence of an SWMI did not affect the pattern of performance on the Hebbian-learning task. Explicit learning of paired associates was better for familiar than for unfamiliar words. A working memory impairment was associated with reduced paired-associate learning generally, whereas a language impairment had an impact only on novel word learning. Overall, explicit learning of familiar words was correlated with domain-general working memory and implicit learning, while the learning of nonwords was additionally related to general intelligence, phonological short-term memory, and language abilities.

Implicit learning of the sequence of pictures of familiar nouns was evident in our Hebb visual–auditory learning task, adding to the growing evidence of Hebbian learning in children (Majerus et al., 2008; Mosse & Jarrold, 2008, 2010). Consistent with Mosse and Jarrold (2008), the Hebbian learning pattern observed in our typically developing group was characterized by maintenance of performance on Hebb trials and by a decrement on filler trials. This pattern contrasts with that of adults, who have shown maintenance of performance on filler trials and a reliable increase in scores on Hebb trials (Hebb, 1961). Mosse and Jarrold (2008) suggested that children’s performance suffers from fatigue or proactive interference effects on the filler lists, and that these effects are reduced on the repeated Hebb list.

Neither working memory nor language impairment resulted in an impairment in implicit sequence learning in the present study. Evidence of intact implicit sequence learning in children with SLI is consistent with Majerus et al.’s (2008) report of a Hebb repetition effect in SLI and Mosse and Jarrold’s (2010) findings of Hebb-list learning in individuals with Down syndrome. Contrary to our expectations, we also observed intact implicit sequence learning in children with SWMI. Although our results have little to say about the nature of the long-term sequence-learning mechanism involved in implicit learning, this ability was not related to domain-general working memory or domain-specific language-learning processes in the present study. It is important to recall, however, that task demands were adjusted in the present work to equate short-term memory demands and to provide a consistent working memory load across participants. As a result, the impact of individual differences in working memory on implicit learning may have been underestimated.

Hebbian learning across visual and auditory modalities was also examined in the present study, as has been demonstrated for adult groups (Page et al., 2006). Findings for our typically developing group provided clear evidence of Hebbian learning in the visual modality; however, we failed to show a reliable Hebb effect for our auditory trials. In fact, we did find maintenance of performance on Hebb trials in the auditory modality, in agreement with our own findings for the visual modality and consistent with work by Mosse and Jarrold (2008, 2010) demonstrating Hebbian learning across spatial and verbal domains in children. Nevertheless, the filler-trial decrement was not observed, due to poor performance on the first half of filler trials in the auditory modality leading to consistent scores on the filler trials across list halves. One possible explanation for these results is that fatigue effects persisted on the filler trials from visual to auditory modalities, an effect that was not seen on the Hebb trials due to carryover learning from the visual to the auditory Hebb list. It may also be that the fewer trials completed in the auditory modality mitigated the effects of fatigue.

Differential Hebbian-learning effects across modalities were noted for SLI but not for SWMI groups in the present study. Lower performance for the SLI group was observed for both the auditory modality and the first half of trials. One motivation for examining learning across modalities was to focus on the ability to recode visual information to phonological forms. Effective recoding would lead to better recall on initial Hebb auditory trials, as was found for the typically developing group. The SLI group, however, performed more poorly in the auditory condition, regardless of list type. These results point to a generalized difficulty with the auditory modality rather than a specific deficit in phonological recoding. Difficulty learning in the auditory modality is a hallmark of SLI (Tallal, 2000) and is in agreement with findings of verbal but not visuospatial deficits in working memory (Archibald & Gathercole, 2007) and learning tasks (Lum et al., 2010). The finding of lower first-half SLI scores is more difficult to interpret. One possible explanation relates to the auditory advantage enjoyed by the typically developing group overall, and on the first half of Hebb lists in particular. It may be that our study was underpowered to reveal a three-way interaction between modality, list, and group that would support this notion.

An explicit learning advantage emerged for associating pairs involving familiar over novel phonological sequences for all groups in the present study. Recall of familiar words is supported by long-term memory (Gathercole, 1995), which facilitates new learning. Associative learning was differentially related to additional measures, depending on whether the associations to be learned involved familiar or novel words. The learning of associations with both familiar and novel words was correlated with working memory, given the needs to both store the phonological form and process the associations in both cases. Explicit word and nonword learning was also related to measures of implicit learning, as has been found in many previous studies (Mosse & Jarrold, 2008; Page & Norris, 2009; Szmalec et al., 2009). The learning of associations with nonwords, however, was additionally related to general intelligence, phonological short-term memory, and language abilities. These findings replicate and extend Mosse and Jarrold’s (2008) report of distinct relationships between new-word learning and both phonological short-term memory and domain-general memory factors. Specifically, our results suggest that phonological short-term memory is linked to novel-word learning in particular, whereas domain-general working memory is related to the learning of both word and nonword associations. The additional results for our impaired groups parallel these findings. SWMI resulted in poor learning of paired associates with both words and nonwords, while the difficulties resulting from SLI were specific to the learning of novel phonological forms. The differences in these patterns are surprisingly marked, and underscore the considerable demands of new-word learning.

While the motivation of this study was to examine factors related to word learning, our results have important clinical implications as well. In particular, our findings provide additional evidence that language-based impairments are separable from working-memory-based impairments. We previously demonstrated dissociable specific language and working memory impairments in children using standardized tests (Archibald & Joanisse, 2009). In the present work, we found that language and working memory impairments have different and specific effects on learning verbal information: Children with SLI struggled to learn novel arbitrary sequences and phonological forms when sequences were presented auditorily, whereas those with SWMI showed reduced immediate recall and explicit learning generally. It would follow from these results that specific intervention approaches and differentiated teaching strategies may be needed to address the disparate needs of each of these groups.

Caveats and limitations

The present study has several limitations. We equated short-term memory demands by adjusting the supraspan level to each participant’s digit span. As a result, some children recalled lists of seven items, and others of four. In recall, the chances of guessing correctly would be greater for shorter than for longer spans, which may have placed the children with longer spans at a disadvantage. However, given that mean span length did not differ systematically between groups, this limitation was unlikely to drive the effects of interest reported here. Another limitation relates to the responses in the visual versus auditory modalities of the Hebb task. All of the pictures were presented in a fixed order on the recall screen for the visual trials, whereas no such cues were available during auditory recall. As such, support available from visuospatial short-term memory would have been greater for the visual than for the auditory trials. Nevertheless, performance on auditory trials was higher overall and was differentially affected by list type and half (i.e., scores were significantly lower on the first half of auditory filler trials generally), suggesting that a general advantage for visual over auditory trials could not account for these results. Future research aimed at systematically varying carryover learning from visual to auditory or auditory to visual modalities would assist in understanding these relationships. Finally, the present study included children with SWMI, a group that lacks confirmatory evidence at present. Further work will be needed to identify and define the characteristics of this group.

Conclusion

Word and sequence learning were examined in children with language and/or working memory impairment and in typically developing children. Language abilities were associated with a specific difficulty learning auditory–verbal information, even when differences in working memory and nonverbal intelligence were taken into account. Working memory was related to explicit learning across modalities generally. These findings suggest that limitations in domain-specific phonological and domain-general memory factors have different and specific detrimental effects on verbal learning.