The Role of Attention, Language Ability, and Language Experience in Children's Artificial Grammar Learning

Abstract

Purpose:

The current study examined the role of attention and language ability in nonverbal rule induction performance in a demographically diverse sample of school-age children.

Method:

The participants included 43 English-speaking monolingual and 65 Spanish–English bilingual children between the ages of 5 and 9 years. Core Language Index standard scores from the Clinical Evaluation of Language Fundamentals–Fourth Edition indexed children's language skills. Rule induction was measured via a visual artificial grammar learning task. Two equally complex finite-state artificial grammars were used. Children learned one grammar in a low attention condition (where children were exposed to symbol sequences with no distractors) and another grammar in a high attention condition (where distractor symbols were presented around the perimeter of the target symbol sequences).

Results:

Overall, performance in the high attention condition was significantly worse than performance in the low attention condition. Children with robust language skills performed significantly better in the high attention condition than children with weaker language skills. Despite group differences in socioeconomic status, English language skills, and nonverbal intelligence, monolingual and bilingual children performed similarly to each other in both conditions.

Conclusion:

The results suggest that the ability to extract rules from visual input is attenuated by the presence of competing visual information and that language ability, but not bilingualism, may influence rule induction.

While most children effortlessly master developmentally appropriate language skills, some children experience protracted or atypical developmental trajectories. Research into individual differences in language ability has yielded several candidate mechanisms underlying variation, impacting theoretical perspectives and clinical intervention approaches to language acquisition. One such mechanism is statistical learning—the process of detecting and extracting patterns in the input without explicit knowledge of the structural regularities learned (e.g., Fiser & Aslin, 2001; Knowlton & Squire, 1996; Saffran, Aslin, & Newport, 1996; Saffran, Newport, & Aslin, 1996). Studies have shown that performance on statistical learning tasks predicts individual differences on a number of linguistic processes in children (e.g., Graf Estes et al., 2007; Mainela-Arnold & Evans, 2014) and adults (e.g., Conway et al., 2010; Misyak & Christiansen, 2012), suggesting that the ability to track statistical regularities may be important for learning linguistic sequential structures (i.e., sounds in words) and distributional structures (i.e., tense agreement). Deficits in statistical learning in developmental language disorder (DLD) have also been cited as evidence for a mechanistic, causal link between statistical learning and language ability (e.g., Lammertink et al., 2017; Obeid et al., 2016).

Similarly, attention—broadly defined as a set of cognitive processes involved in the selection of stimuli for processing (e.g., Gomes et al., 2000)—is theorized to underlie the acquisition of language and has been linked to individual differences in language processing abilities in typically developing (e.g., de Diego-Balaguer et al., 2016; Kannass & Oakes, 2008; Yu et al., 2019) and language-impaired populations (e.g., Kapa & Plante, 2015; Kapa et al., 2017). The evidence for the role of attention in statistical learning, however, is not as robust. Prior studies of statistical learning and attention, which have been largely conducted in monolinguals, have yielded mixed findings, with some studies finding that increasing attentional load interferes with statistical learning (e.g., Toro et al., 2005, 2011), and some studies finding that manipulations of attentional load do not shape statistical learning outcomes (e.g., Musz et al., 2015; Saffran et al., 1997). Notably, the literature linking statistical learning and language skills and the literature linking statistical learning and attention rarely intersect, despite the fact that two-way connections among the three constructs have been extensively studied. No less notably, studies testing the role of attention in statistical learning as well as studies linking statistical learning and language skills have largely been limited to monolingual populations, despite the extensive literature that considers the interactions between bilingual experience and attention (e.g., Adesope et al., 2010; Blom et al., 2017), as well as between bilingual experience and language outcomes (e.g., Hurtado et al., 2014). In this study, we extend the existing literature by considering the role of attention in the relationship between statistical learning and language ability within a diverse sample of monolingual and bilingual children with poor-to-above-average language skills.

Statistical Learning and Attention

Statistical learning is defined as extraction of co-occurrence statistics in the input without trying to learn and without becoming aware that learning occurred (e.g., Fiser & Aslin, 2001; Knowlton & Squire, 1996; Saffran, Newport, & Aslin, 1996). Statistical learning has been proposed to be a domain-general mechanism and has been implicated in the development of both nonlinguistic (e.g., Kirkham et al., 2002; Liao & Masters, 2001) and linguistic (e.g., Onnis et al., 2003; Romberg & Saffran, 2010) skills. Learning via incidental exposure has been observed in as little as 2 min (e.g., Saffran, Aslin, & Newport, 1996) and has been documented across the life span from infancy (e.g., Bulf et al., 2011; Hay et al., 2011) into adulthood (e.g., Amso & Davidow, 2012).

The degree to which attention can modulate statistical learning performance remains unclear. The role of attention in statistical learning has been examined most frequently using divided-attention tasks, yielding conflicting evidence. On the one hand, studies have shown that infants (e.g., Saffran, Aslin, & Newport, 1996), children, and adults (e.g., López-Barroso et al., 2016; Musz et al., 2015; Saffran et al., 1997) can learn adjacent transitional probabilities (i.e., relationships between syllables in a continuous sequence that make a word-like unit) and nonadjacent transitional probabilities (i.e., relationships between nonadjacent syllables) without the support of attention. For example, across four experiments, Musz et al. (2015) observed that adult participants could extract temporal regularities for both unattended and attended stimuli sets in a visual statistical learning paradigm when engaged in a shape detection task. In their study, changing the cover task (i.e., shape repetition detection task; Experiment 2) and the complexity of the statistical regularities (Experiment 3) did not alter the findings.

Conversely, other works have revealed effects of attention in statistical learning (e.g., Campbell et al., 2012; Toro et al., 2005, 2011; Turk-Browne et al., 2005). For example, in a series of experiments, Toro et al. (2005) examined word segmentation performance under low and high attention demand conditions. Participants in the low attention condition listened passively to an artificial speech stream while concurrently being exposed to noises of common objects (i.e., bell ring; Experiment 1), pictures (Experiment 2), and pitch changes (Experiment 3). Participants in the high attention condition were also exposed to the artificial speech stream but were instructed to press a button each time they detected a repetition in the nontarget stream. Results demonstrated that word segmentation was significantly compromised when both auditory and visual attention were diverted away from the target speech stream. In a later study, Toro et al. (2011) examined the effects of attention in word segmentation as well as in generalization of rules governing target words following the design used in the work of Toro et al. (2005). Results over four experiments indicated that divided attention attenuated both word segmentation performance and generalization of nonadjacent regularities, but not adjacent regularities, suggesting that the degree to which attention modulates statistical learning may vary depending on the type of rule structure.

Findings from the work of Toro et al. (2011) are largely consistent with the view that statistical learning may involve both attention-independent processes and attention-dependent processes (Batterink et al., 2015; Conway, 2020). Under this account, attention-independent processes are always “on,” and attentional resources are recruited in the learning of more complex regularities (Conway, 2020). For example, the learning of simple adjacent regularities in word segmentation tasks, which has largely been the focus of prior work examining contributions of attention to statistical learning (e.g., Campbell et al., 2012; Musz et al., 2015; Saffran et al., 1997; Toro et al., 2005, 2011), may rely more on implicit processes and is thus acquired automatically and effortlessly (Conway, 2020). Therefore, word segmentation may not be demanding enough to consistently yield effects of divided attention. In contrast, learning complex distributional regularities and abstract rule-based patterns may require the involvement of attentional resources (Conway, 2020; Toro et al., 2011). It remains an open question, however, whether different types of statistical learning paradigms are variably susceptible to attention manipulations.

In this study, we test the role of attention in statistical learning in an artificial grammar learning (AGL) task. AGL is a type of rule induction paradigm that arguably involves a much more challenging process than a word segmentation task, in that it entails generalization of distributional regularities extracted during exposure to novel stimuli. In an AGL task, participants are exposed to a series of symbol sequences that are generated by a system of rule-based patterns (i.e., a grammar). In the exposure phase, participants are exposed to grammatical symbol sequences that typically vary in length from three to six symbols. After passive exposure, participants engage in a grammaticality judgement task where they are instructed to identify whether novel symbol sequences are grammatical (i.e., follows the rule-based pattern) or ungrammatical (i.e., violates the rule-based pattern). Studies have shown that children (e.g., Pavlidou & Williams, 2014; Pavlidou et al., 2009) and adults (e.g., Schiff et al., 2017; Westphal-Fitch et al., 2018) develop implicit knowledge of the grammatical structure following passive exposure and are sensitive to novel grammatical and ungrammatical exemplars.

The role of attention in AGL performance has been examined in only two studies, both conducted with adults (Hendricks et al., 2013; Tanaka et al., 2008). Unlike findings in the work of Musz et al. (2015) who observed learning of both unattended and attended stimuli sets in a word segmentation task, Tanaka et al. (2008) found that adults extracted rules for attended sequences only. Adults did not learn the rules for unattended sequences that were presented simultaneously. However, Hendricks et al. (2013) observed similar AGL performance in adults who were engaged in a working memory dual task and a single task during the learning phase (Experiment 1). Deleterious effects of divided attention on AGL performance were only observed when participants engaged in a dual task at test (Experiment 1) and when perceptual cues were changed between the learning phase and test phase (Experiment 2). Contrasting findings among these studies are consistent with Conway's (2020) theorizing that complex rule abstraction (e.g., as tested in the work of Hendricks et al., 2013; Tanaka et al., 2008) may require greater attentional resources than the learning of simple adjacent probabilities (e.g., Musz et al., 2015). In addition to the role of input complexity, it is likely that the variety of attentional manipulations (e.g., concurrent load, dual task, divided attention via distractor stimuli) in different modalities (e.g., auditory vs. visual) across studies contributes to the mixed findings in literature. Differences in methodologies and approaches make it difficult to determine why some manipulations incur performance costs while others do not.

In the current study, we tested the role of attention via a divided attention task following a number of prior studies (e.g., Hendricks et al., 2013; Musz et al., 2015; Toro et al., 2005, 2011). To increase attentional demands, we positioned distractor stimuli around the perimeter of target symbol sequences during learning. Both target stimuli and distractor stimuli were presented in the same modality; both were visual symbols. The use of nonverbal visual symbols allowed us to keep attentional demands constrained to one domain while simultaneously controlling for differences in children's language exposure and experience. Additionally, while the grammars used were designed to mimic sentence construction in natural languages, the grammatical regularities generated by each grammar were not language specific. The symbol sequences did not follow, for example, English-specific sentence constructions, leveling the learning load for children with diverse language learning histories.

To our knowledge, ours is the first study to test whether imposing a higher attention load would reduce performance on an AGL task in children. Children may be especially sensitive to attentional manipulations compared to adults. We hypothesized that if attention gates rule induction, then learning in the presence of irrelevant stimuli will negatively impact AGL performance in children. If rule induction operates without the need of attention, then children will learn grammatical structures equally well in the presence and absence of irrelevant stimuli. In framing this question, we considered whether children's language ability and language learning experience would influence rule induction under different attentional loads.

Statistical Learning and Language Ability

Statistical learning has been implicated in the development of linguistic skills (e.g., Onnis et al., 2003; Romberg & Saffran, 2010), such as speech segmentation (e.g., Saffran, 2003), identifying syntactic phrase boundaries (e.g., Saffran, 2002), integrating prosodic and morphological cues in the learning of phrase structure (e.g., Morgan et al., 1987), detecting long-distance relationships between words (e.g., Onnis et al., 2003), and detecting orthographic regularities of written words (e.g., Pacton et al., 2001). Additionally, performance on statistical learning tasks has been shown to predict individual differences in children's phonological processing skills (e.g., Mainela-Arnold & Evans, 2014), word learning skills (e.g., Graf Estes et al., 2007; Singh et al., 2012), syntactic comprehension (e.g., Kidd & Arciuli, 2016), and literacy skills (e.g., Arciuli & Simpson, 2012; Elleman et al., 2019).

A growing area of research has focused on investigating how individual differences in cognitive abilities (i.e., attention, memory, language) in typically developing populations support and constrain statistical learning mechanisms (e.g., Conway et al., 2010; Misyak & Christiansen, 2012; Scott & Fisher, 2012; Vlach & DeBrock, 2017). For example, Vlach and DeBrock (2017) examined whether age, visual and auditory recognition memory, and receptive vocabulary skills predicted children's statistical word learning performance. The results revealed that children's vocabulary skills and visual recognition memory predicted their statistical word learning performance over and above the effects of age. Similarly, Scott and Fisher (2012) observed that only children with robust vocabularies learned novel transitive verbs in a cross situational word learning paradigm; children with poorer vocabularies did not demonstrate evidence of learning novel transitive verbs.

Moreover, studies have shown that individuals with DLD perform poorer on statistical learning tasks compared to typically developing children and neurotypical adults; persons with DLD are less sensitive to both adjacent and nonadjacent statistical regularities (e.g., Obeid et al., 2016; Plante et al., 2002). While none of the children in our study had a formal diagnosis of DLD, a number of them scored poorly on our standardized language assessment and had a history of risk factors typically associated with DLD (i.e., recurrent ear infections; low mother's years of education). We hypothesized that if AGL is sensitive to weakness in language skills within the normal range of language ability (i.e., in the absence of a formal DLD diagnoses), then children with weak language skills may show broad weaknesses in learning distributional regularities in the present experiment, independent of an attentional manipulation. However, if AGL performance is not sensitive to variations in language ability within the normal range, then children with weak language skills will learn distributional regularities equally well to children with strong language skills.

We also had specific hypotheses with respect to how variation in language skills would interact with an attentional manipulation. Children with weak language skills may demonstrate subtle weaknesses in selective attention skills and have difficulties ignoring irrelevant learning cues as seen in children with DLD (e.g., Kapa et al., 2017; Marton, 2008; Marton et al., 2014; Pauls & Archibald, 2016; Spaulding, 2010). It is plausible that attending to irrelevant stimuli may reduce the amount of exposure to target statistical regularities, negatively impacting the ability to learn rule-based patterns in the input. Therefore, we hypothesized that if rule induction depends on attentional processes, then AGL performance in the presence of distracting stimuli would be disproportionately challenging for children with weaker language skills compared to children with robust language abilities.

A strength of our study is that we examined the effects of language ability as a continuum in a sample of children with poor-to-above-average language scores. Examining language as a continuum versus categorically may be a more robust analytical strategy given the lack of agreement about DLD diagnostic criteria (Bishop, 2017) and the resulting differences in identification criteria across different settings. Inconsistencies in identification criteria for language impairment are especially problematic in language-minority bilingual children, leading both to over- and underidentification of language impairment in this population (e.g., Morgan et al., 2015; Samson & Lesaux, 2009). By including bilingual children with a range of language skills, we were able to test whether language ability, bilingualism, or both influence AGL performance through attention.

Bilingualism

The effects of bilingualism on the development of attention continue to be heavily contested. Some researchers have failed to find reliable differences between monolinguals and bilinguals (e.g., Duñabeitia & Carreiras, 2015, Gathercole et al., 2014; Paap & Greenberg, 2013; Paap et al., 2015), while others find evidence for attentional advantages in behavioral studies and find support in neuroimaging data showing meaningful neuroanatomical changes in the brains of bilingual language learners (e.g., Adesope et al., 2010; Blom et al., 2017; Costa et al., 2008; Crivello et al., 2016; Hosoda et al., 2013; Schlegel et al., 2012). When observed, bilingual advantages have been most consistently found in children and older adults (e.g., Barac et al., 2016; Bialystok, 2011; Bialystok et al., 2009; Kapa & Colombo, 2013).

Studies examining the effects of bilingualism in statistical learning have yielded similarly complex findings. Some studies have observed broad bilingual advantages in statistical learning performance (e.g., Antovich & Graf Estes, 2018; Benitez et al., 2016; de Bree et al., 2017; Escudero et al., 2016; Kovács & Mehler, 2009; Onnis et al., 2018; Poepsel & Weiss, 2016), while others have not (e.g., Bulgarelli et al., 2019; Poepsel & Weiss, 2016; Yim & Rudoy, 2013).

Critically, only a few studies have examined the role of bilingual experience in statistical learning through a divided-attention paradigm in order to test the possibility that bilingualism influences statistical learning through attention (e.g., Bartolotti et al., 2011; Kovács & Mehler, 2009; Poepsel & Weiss, 2016). In these studies, participants were required to selectively attend to one of two conflicting segmentation cues to successfully segment words. Bilingual infants (Kovács & Mehler, 2009) and adults (Onnis et al., 2018; Poepsel & Weiss, 2016) were more flexible learners of multiple structural regularities than monolinguals. However, these bilingual advantages have not consistently been observed in adults (e.g., Bartolotti et al., 2011). In their study, Bartolotti et al. (2011) found that for young adults, inhibitory control skills, and not bilingual experience, were associated with statistical learning in a high attention condition on a word segmentation task. Therefore, it is unclear whether bilingual experience would enhance children's statistical learning performance, in the presence (or the absence) of interference. While overall group differences in rule induction were possible, we were more interested in whether attention influenced rule induction differently in bilingual children versus monolingual children. We hypothesized that if bilingual language experience enhances statistical learning performance broadly, then bilingual children will outperform monolingual children on the AGL task across both conditions. Conversely, if effects of bilingualism are specific to conditions of increased attentional demand, as indicted in previous studies linking enhancements in attention to bilingual language experience (e.g., Adesope et al., 2010; Blom et al., 2017), then effects of bilingualism would be observed only in the high attention condition.

The Current Study

Research on the role of attention in statistical learning suggests that statistical learning may be compromised when learning takes place in the presence of competing information. The first question asked in this study is whether this is the case for rule induction, a challenging statistical learning task that requires generalization of newly learned grammatical regularities. The second question asked in this study is whether language ability and bilingual experience contribute to rule induction performance under different attentional demands in children. To test these research questions, we administered an AGL task to index rule induction performance under light and heavy attention demands to monolingual and bilingual children spanning a wide range of language ability.

Method

Participants

This study was reviewed and approved by the Education and Social/Behavioral Science Institutional Review Board at University of Wisconsin–Madison. Participants' legal guardian or next of kin provided informed consent, and children provided oral assent. One hundred twenty-six children between the ages of 5 and 9 years were recruited from the Madison, WI, metropolitan area. This age range was selected because children show developmental improvements in statistical learning abilities (e.g., Arciuli & Simpson, 2011) and inhibitory control processes (i.e., the ability to ignore distracting information; Ikeda et al., 2014) within this age range, allowing us to capture variability in AGL performance under light and heavy attention conditions. Participants were excluded due to neurodevelopmental disability or confirmed language disorder (n = 2), response bias scores of 1 or −1 in experimental task (i.e., pressing one button for all test items; n = 6), and missing data (n = 10). This study included 108 children: 43 English-speaking monolinguals (23 male, 20 female; M _age = 6.88 years; SD _age = 1.26) and 65 Spanish–English bilinguals (31 male, 34 female; M _age = 7.38 years; SD _age = 1.31). None of the parents expressed concerns for any of the children. Additionally, none of the children were receiving speech-language therapy services or had referrals in process for language, behavior, or academic concerns at the time of testing. Children were considered monolingual if they had less than 5% consistent exposure to any language other than English at any time. Children were considered bilingual if they were exposed to Spanish at least 20% of their waking hours. Information about bilingual children's current language exposure, language dominance, and language preference was collected through parent questionnaires. We also collected information about primary caregivers' language use, language proficiency, and primary caregivers' level of education. All children passed a bilateral pure-tone hearing screening at 25 dB at 1000, 2000, and 4000 Hz. Parents confirmed children had normal or corrected vision acuity. See Table 1 for monolingual and bilingual group characteristics.

Table 1.

Participant characteristics.

Variables	Monolinguals	Bilinguals	t
	M (SD)	M (SD)
n	43 (23 boys)	65 (31 boys)
Age (years)	6.89 (1.24)	7.38 (1.31)	1.99*
Primary caregiver's years of education	17.29 (2.66)	13.60 (4.24)	−5.56**
Nonverbal IQ a	111.67 (18.05)	100.82 (14.2)	−3.32***
CELF-4 Core Language Index b	112.58 (12.69)	88.57 (18.63)	−7.97***
CELF-4 Spanish Core Language Index c	—	88.85 14.95
Language Ability d	112.58 12.69	97.12 13.47	6.05***
Age at first exposure to English (months)		13.28 (18.87)
Age at first exposure to Spanish (months)		11.85 (23.47)
Current English exposure e (%)		67.10 (28.19)
		n
Language heard at school
English only		23
Spanish at least 50% of time		42
Language heard at home
Mostly English		19
Mostly Spanish		36
Both English and Spanish		10
Language spoken at home f
Mostly English		23
Mostly Spanish		30
Both English and Spanish		10

^a Matrices subtest of the Kaufman Brief Intelligence Test–Second Edition.

^b Standard scores of subtests of the Clinical Evaluation of Language Fundamentals–Fourth Edition (CELF-4).

^c Standard scores of subtests of the CELF-4 Spanish.

^d Highest Core Language Index score from either CELF-4 English or Spanish.

^e Parental report of exposure to language during waking hours in a typical week.

^f Two parents failed to report the languages spoken at home.

^* p < .05.

^** p < .01.

^*** p < .001.

Standardized Measures

The Clinical Evaluation of Language Fundamentals–Fourth Edition (CELF-4; Semel et al., 2003) was used to evaluate each participant's expressive and receptive language abilities in English. The CELF-4 Spanish (Wiig et al., 2006) was used to evaluate bilingual participants' expressive and receptive language abilities in Spanish.

The Visual Matrices subtest of the Kaufman Brief Intelligence Test–Second Edition (Kaufman & Kaufman, 2004) was used to assess each participant's nonverbal intelligence. Items on the subtest require understanding of spatial relationships, use of abstract reasoning, and use of problem-solving strategies for both meaningful (i.e., animals and plants) and abstract (i.e., shapes and figures) items.

Language Ability

The CELF-4 Core Language Index standard score, which measures overall receptive and expressive language skills, was used to index language ability. For bilingual children, language ability was indexed by children's highest Core Language Index score from either CELF-4 English or CELF-4 Spanish (Crespo et al., 2019). We used the child's best Core Language Index score in an effort to not penalize bilingual children for length of language experience, unbalanced language exposure, and exposure-based low language scores in the nondominant language.

Core Language Index standard scores from the CELF-4 were used to index language ability for all monolingual children and 27 bilingual children; scores from the CELF-4 Spanish Edition were used to index language ability for 38 bilingual children. In the process of testing, we identified one monolingual child and 11 bilingual children with poor language skills who obtained standard scores below 1.25 SDs (standard scores ≤ 85). We included these children in the analyses because there was no language disorder diagnosis or parent concerns reported at the time of testing. Participants' highest CELF-4 Core Language Index standard scores ranged from 66 to 134.

AGL Task

Two equally complex finite-state artificial grammars devised by Reber and Allen (1978) were used to generate grammatical and ungrammatical symbol sequences. Each grammar generated the same number of symbol sequences because the mathematical structural functions that characterize finite-state systems are identical in both grammars (Reber & Allen, 1978). To avoid language confounds, five easily discernible symbols from the Wingdings font from Microsoft Word 2003 replaced the original letter sequences. Symbols were black in color and were presented in a 60-point font (see Figures 1a to 1b for the visual representation of the two grammars). Symbols appeared simultaneously and were arranged horizontally in a row in the middle of the screen. See the Appendix for a list of symbol sequences for Grammar 1 and Grammar 2.

Figure 1.

(a). Grammar 1. (b). Grammar 2. This grammar generates symbol sequences that are three to six symbols long. Grammatical symbol sequences were generated by beginning at Node 1 and following the directionality of the grammar as indicated by the arrows to a terminating node (Node 5 or 6). Circular arrows indicate that symbols in that position can be repeated or skipped. Ungrammatical symbol sequences were generated by systematically changing symbol positions in a grammatical sequence. An example of an ungrammatical sequence for both grammars would follow the node sequence 3555. In this example, the symbol sequence did not start at Node 1.

Grammatical Sequences

Grammatical sequences were generated by starting at the first node and following the directionality of the grammar to an exit node (Reber & Allen, 1978). Each grammar generates five grammatical sequences with three symbols, seven grammatical sequences with four symbols, 11 grammatical sequences with five symbols, and 18 grammatical sequences with six symbols. Each grammar generates 41 grammatical symbol sequences in total. Nineteen grammatical sequences were used during the training phase, and 19 different grammatical sequences were used during the testing phase.

Ungrammatical Sequences

Ungrammatical sequences were generated by systematically introducing violations to symbol positions in a grammatical sequence (Reber & Allen, 1978). For example, in Grammar 1, only two symbols can legally start a sequence. A violation in symbol position can be introduced by generating a sequence with any other symbol in the first position, rendering the sequence ungrammatical.

Reber and Allen (1978) originally generated 25 ungrammatical sequences by systematically introducing violations to symbol positions in 22 sequences and using three grammatical sequences spelled backward. Violations were made by inserting incorrect symbols in certain positions in the sequences. These symbols would not have appeared in these positions if the rules governing the order of symbol sequences for each grammar were followed. Incorrect symbols that appeared in ungrammatical sequences were chosen at random from the set of symbols used in each respective grammar. For the current study, 19 ungrammatical sequences were generated by introducing the following violations: Five sequences had a violation in the first position, five sequences had a violation in the second position, four sequences had a violation in an internal position, and five sequences terminated incorrectly.

Conditions

The design of each grammar was manipulated such that each grammar could be presented in a high attention demand condition and a low attention demand condition. The high attention condition was created by introducing two distracter symbols around the perimeter of the target sequence during each trial in the exposure phase. Distractor symbols consisted of five black, 60-point font symbols from the Wingdings font from Microsoft Word (2003) that were distinct from the symbols used to construct Grammars 1 and 2. Distractor symbols did not follow a rule-based pattern. Instead, they appeared randomly in any one of the following six positions: top left corner, top center, top right corner, bottom left corner, bottom center, and bottom right corner. Positions of the distracter symbols were randomized for each trial. The low attention condition consisted of the black symbol sequences presented without the competing distractor symbols around the perimeter of the target sequence. See Figures 2a to 2b for example trials in the low and high attention conditions for each grammar.

Figure 2.

(a). Low attention condition example trials for Grammar 1 and Grammar 2. (b). High attention condition example trials for Grammar 1 and Grammar 2. In the high attention condition, black symbol sequences were presented with competing distractor symbols around the perimeter of the target sequence. Distractor symbols did not follow a rule-based pattern. Instead, they appeared randomly in any one of the following six positions: top left corner, top center, top right corner, bottom left corner, bottom center, and bottom right corner. Positions of the distracter symbols were randomized for each trial.

Procedure

Participants completed standardized language measures as well as experimental tasks over the course of three sessions. Participants were tested individually in a quiet room at the Waisman Center. Each participant completed an exposure and a testing phase in one of two conditions during two different sessions on two different days. For example, if a participant completed Grammar 1 in the low attention condition in the first session, the participant completed Grammar 2 in the high attention condition in the second session. Condition, grammar, and order were counterbalanced across participants.

Before the exposure phase, participants were presented with the target shapes aligned in a circle in random order and were told to watch a movie with these shapes. No further instructions were given regarding finding a rule-based pattern or where to allocate attention during the movie. During the exposure phase, participants were shown 19 grammatical sequences repeated in four blocks. Symbol sequence order was randomized in each block. Each sequence was presented for 2,000 ms followed by a fixation cross for 500 ms and an interstimulus interval of 500 ms. To encourage the children to look at the monitor, a smiley face was presented between each block along with praise from the experimenter and a verbal prompt to look at all the shapes.

Immediately following the exposure phase, participants were introduced to a grammaticality judgment task where they were presented with 19 novel grammatical sequences and 19 ungrammatical sequences. Before the grammatically judgment task, participants were told that all the symbols presented at exposure followed a “secret pattern.” The participants were encouraged to make their best guess as to whether each novel sequence in the grammatically judgment task followed the secret pattern. The order of presentation of grammatical and ungrammatical symbol sequences was randomized for each participant. Each sequence was presented for 4,000 ms followed by a fixation cross for 500 ms and an interstimulus interval of 500 ms. Participants had 4,000 ms to make a grammaticality judgment. If the participant did not make a decision, the trial terminated and moved to the next sequence.

Analyses

Fifty trials with reaction times below 150 ms were removed. These trials were test responses to the grammatically judgement task. A total of 7,541 observations were included in a logistic mixed-effects model constructed in RStudio, Version 1.2.5001 (RStudio Team, 2019) using the lme4 package (Bates et al., 2015). Children's accuracy data were analyzed to examine the extent to which predictors increased or decreased the likelihood (log-odds) of making an accurate response. Fixed effects included an interaction between group (contrasted coded [−.5, .5], monolinguals vs. bilinguals), condition (contrasted coded [−.5, .5], low attention vs. high attention), and language ability (mean centered); all lower ordered two-way interactions and main effects were included. Due to the sample's wide age range and group differences, age, primary caregiver's years of education, and nonverbal IQ were included as covariates. The most complex model was fitted following Barr et al.'s (2013) suggested “keep it maximal” approach. By-item random slopes for the interactions and main effects of group, condition, and language ability were removed in a stepwise model comparisons approach to resolve convergence and singularity issues (Brauer & Curtin, 2018). The final model included a by-participant random intercept, a by-participant random slope for the effect of condition, and a by-item random intercept. The by-participant random intercept and by-participant random slope for the effect of condition were uncorrelated to resolve singularity issues. The inclusion of by-item random slopes for the effects of group and/or language ability did not alter the pattern of significant findings. Additionally, the effects of condition order (completing low vs. high condition first) and grammar (Grammar 1 vs. Grammar 2) were examined, and the results suggest that order and grammar did not significantly predict AGL performance (ps > .05) or interact with predictor variables (ps > .05).

Results

All participants contributed data in the low attention condition, but one monolingual and five bilingual participants did not contribute data in the high attention condition. These participants did not contribute data due to failure to return for the final session (n = 4) and a computer error (n = 2). Collapsing across group, results from t tests revealed that overall, children learned the grammar above chance (i.e., 0.50) in the low attention condition (M = 0.54, SD = 0.07; range: 0.41–0.75; t(107) = 5.61, p < .001), but not in the high attention condition (M = 0.51, SD = 0.09; range: 0.19–0.72; t(101) = 0.82, p = .41). Monolingual children performed above chance in the low attention condition (M = 0.54, SD = 0.07; range: 0.41–0.71; t(42) = 3.69, p < .001), but not in the high attention condition (M = 0.51, SD = 0.09; range: 0.29–0.72; t(41) = 0.59, p = .56). Similarly, bilingual children also performed above chance in the low attention condition (M = 0.54, SD = 0.07; range: 0.42–0.75; t(64) = 4.20, p < .001), but not in the high attention condition (M = 0.51, SD = 0.09; range: 0.19–0.71; t(59) = 0.57, p = .57).

Model results revealed a significant main effect of condition, (B = −0.13, SE = 0.06, z = −2.26, p = .02), such that children were significantly more likely to be accurate in the low attention condition than in the high attention condition (OR = .88, 95% CI [0.79, 0.98]). A significant Condition × Language Ability interaction was also observed, (B = 0.12, SE = 0.06, z = 2.11, p = .04; OR = 1.12, 95% CI [1.01, 1.25]; see Figure 3). To interpret the significant interaction, the simple effects of language ability were tested at each level of condition via a logistic regression model using the generalized linear model function. Language ability did not significantly predict children's accuracy in the low attention condition (z = 0.28, p = .78). However, children with robust language scores were more likely to be accurate in the high attention condition than children with poorer language scores (B = 0.10, SE = 0.04, z = 2.12, p = .03; OR = 1.10, 95% CI [1.01, 1.20]). All other main effects and interactions, including our hypothesized three-way interaction between attention, language ability, and bilingualism, were not significant. See Table 2 for full model results.

Figure 3.

Interaction between condition and language ability. Fitted model values for the Condition × Language Ability interaction term in the full model. AGL = artificial grammar learning; CELF = Clinical Evaluation of Language Fundamentals.

Table 2.

Full model results.

Variables	Full model		Simple effects
			Low attention condition		High attention condition
	B (SE)	z	B (SE)	z	B (SE)	z
Intercept	0.12 (0.22)	0.52	0.13 (0.31)	0.41	0.08 (0.32)	0.26
Age	0.01 (0.02)	0.58	0.03 (0.03)	1.27	−0.01 (0.03)	−0.45
Primary caregiver's years of education	−0.003 (0.01)	−0.50	−0.02 (0.01)	−1.73	0.01 (0.01)	1.09
Nonverbal IQ	−0.001 (0.002)	−0.43	0.0002 (0.002)	0.12	−0.001 (0.002)	−0.63
Condition	−0.13 (0.06)	−2.26*	—	—	—	—
Group	0.02 (0.06)	0.41	−0.08 (0.08)	−0.92	0.12 (0.08)	1.49
Language ability	0.06 (0.03)	1.75	0.01 (0.04)	0.28	0.10 (0.04)	2.12**
Condition × Group	0.13 (0.11)	1.16	—	—	—	—
Condition × Language Ability	0.12 (0.06)	2.11*	—	—	—	—
Group × Language Ability	−0.04 (0.06)	−0.64	−0.04 (0.08)	−0.51	−0.03 (0.08)	−0.35
Condition × Group × Language Ability	0.12 (0.22)	0.22	—	—	—	—
Observations	7,541		3,868		3,673
N	108		108		108
Akaike inf. crit.	10,423.39		5,351.04		5,096.63
Bayesian inf. crit.	10,520.38		5,394.87		5,140.09

Note. Em dashes indicate data not available. SE = standard error; inf. = information; crit. = criterion.

^* p < .05.

^** p < .01.

Discussion

In this study, we examined the effects of attentional load on rule induction in a diverse sample of monolingual and bilingual children. A particular strength of our design was that our AGL paradigm utilized nonlinguistic symbol sequences that followed rule-based patterns that were not specific to English sentence constructions. This allowed us to level possible influences of prior experience for children with different language abilities and language experiences. Additionally, we used two distinct grammars that were matched on complexity to compare performance in low and high attention conditions within participants, limiting learning and carryover effects between conditions. Results revealed that rule induction performance was influenced by the need to ignore irrelevant information in the high attention condition and that children with robust language abilities outperformed children with poorer language skills in the high attention condition. We found no evidence that bilingual language experience impacted children's rule induction or influenced the relationship between attentional load and language ability.

Rule Induction May Be Influenced By Attention

In our study, children showed a marked decrease in performance in the high attention demand condition, suggesting that distractor symbols did interfere with their learning and that distributed attention negatively impacted their rule induction in the visual domain. These results are consistent with an emerging body of literature that suggests that attention may modulate statistical learning performance (e.g., Bartolotti et al., 2011; Jiang & Chun, 2001; Poepsel & Weiss, 2016; Toro et al., 2005; Turk-Browne et al., 2005). Studies have suggested that attention effects on statistical learning are observed to varying degrees depending on the interference manipulation (e.g., divided attention task vs. dual-task paradigm; e.g., Toro et al., 2005). Lack of attention effects in other studies may have been due to the implementation of a dual-task paradigm, a manipulation that has been argued to have little impact on statistical learning (e.g., Hendricks et al., 2013; Toro et al., 2005). Additionally, it is possible that rule induction is more sensitive to attentional manipulations than word segmentation. Compared to learning transitional probabilities, which has been the focus in prior work (e.g., Campbell et al., 2012; Musz et al., 2015; Saffran et al., 1997; Toro et al., 2005, 2011), learning and generalizing distributional probabilities may be more complex and therefore require more attentional resources (Conway, 2020).

Our choice of tasks likely contributed to the pattern of findings we observed, and a different statistical learning paradigm and/or attention manipulation may have yielded a different result. Identifying which task-specific parameters do and do not yield effects of attention in statistical learning is important and would further our understanding of the roles of automatic-implicit processes and non–automatic attention-dependent processes in statistical learning. While we did not design the current study to address this issue, our findings suggest that a divided attention paradigm can yield an effect of attention on AGL performance in children. Future research is needed to examine how rule structure, attentional manipulations, and learning paradigms interact to modulate statistical learning performance. In addition, further research testing participants who are both younger and older than the participants tested here is needed to elucidate how developmental changes in statistical learning abilities and cognitive abilities (i.e., attention and language skills) impact statistical learning performance under different attentional loads.

Effects of Language Ability on Rule Induction

No main effect of language ability was observed; however, language ability was found to moderate the effect of attention in children's rule induction. While children's rule induction performance in the low attention condition did not depend on their language competence, children with robust language ability outperformed children with weaker language skills in the high attention condition. A lack of a significant main effect of language ability is in contrast with the growing body of evidence suggesting that individual differences in language abilities in typically developing populations are associated with statistical learning performance (e.g., Misyak & Christiansen, 2012; Scott & Fisher, 2012; Vlach & DeBrock, 2017). Our measure of language ability (i.e., CELF Core Language Index score) indexes both receptive and expressive language skills across multiple domains of language, and it is likely that such a broad language metric is only distally related to a specific skill like nonlinguistic rule induction. Prior work examining individual differences in statistical learning in typically developing children has linked vocabulary knowledge and cross-situational word-learning performance, a skill arguably more proximal to vocabulary acquisition (e.g., Scott & Fisher, 2012; Vlach & DeBrock, 2017). It is possible that a more specific measure of language ability would have been related to children's performance in the low attention condition in our study as well.

Utilizing nonlinguistic stimuli may have also impacted our pattern of results. It is possible that performance in the low attention condition could have been more sensitive to differences in language ability had we required participants to extract linguistic statistical structure. The use of lexical phonological distributional dependencies used in prior work (e.g., Grunow et al., 2006; Plante et al., 2002) may have yielded differences in rule induction between children with stronger language skills and children with weaker language skills. It is also possible that our study may have required a greater number of children with poor language skills to detect differences in nonverbal rule induction under lighter attentional loads. In this study, only 12 out of 108 children obtained below-average-to-poor language scores (standard score range: 66–82). The inclusion of children with a diagnosis of developmental language disorder in future studies will increase the number of children on the lower end of the language continuum and may help to determine whether weaknesses in language ability are associated with weaknesses in statistical learning under lighter attentional loads. We acknowledge that our hypotheses largely stem from the premise that language abilities may constrain statistical learning. However, it is also a possibility that children's statistical learning abilities have led to their current level of language ability. Therefore, future research is also needed to determine how weaknesses in one or both domains interact and gate learning over time.

Critically, differences in rule induction were observed when attentional demands were heightened such that children with weaker language skills were disproportionately impacted by the presence of distractor stimuli. These results suggest that typically developing children with weaker language skills have trouble discovering statistical dependencies when distracted, possibly due to subtle attention difficulties. Participants with particularly strong language skills (CELF-4 scores < 120) appeared to successfully learn the statistical dependencies (i.e., perform above chance) in the high attention condition, whereas participants with average or below average language skills were at or below chance. Typical of most formal standardized tests, the CELF-4 arguably requires a moderate degree of attention for successful performance. In our study, it appeared that standardized language scores were particularly well suited to capture meaningful variations in attention skills within the typical range of language ability. Future directions should include an independent measure of attention skills and an examination of the effects of different attentional load manipulations in children along the spectrum of attentional abilities. It would also be fruitful to consider whether systematic manipulations of attention within language tasks could be harnessed for clinical diagnostic purposes, with heightened sensitivity to such manipulations characterizing children with weaknesses in language, which would otherwise go undetected, as for instance often is the case with bilingual children.

Effects of Bilingualism on Rule Induction

In this study, we did not find evidence to suggest that rule induction performance was related to language experience. Similar null results of bilingualism have been reported in previous studies where group differences were not observed on a visual and an auditory statistical learning task when no additional demands on attention were imposed (e.g., Yim & Rudoy, 2013). We hypothesized that the high attention condition would yield superior performance in bilinguals and that bilingual language experience would moderate the relationship between language ability and attention, based on evidence suggesting that bilingualism can facilitate performance on attention control (e.g., Barac et al., 2016; Bialystok, 2011; Bialystok et al., 2009; Kapa & Colombo, 2013). However, our findings are congruent with recent work that has failed to find bilingual advantages on tasks of attention control in children (e.g., Duñabeitia et al., 2014; Gathercole et al., 2014), although it is important to note that we did not explicitly probe for the effects of bilingualism on attention control. Future work will need to carefully consider the effects of bilingualism on both statistical learning and attention control, ideally in samples of demographically matched bilingual and monolinguals.

In our study, we controlled for significant group differences in primary caregiver's years of education and nonverbal IQ scores and found no differences in AGL performance between monolinguals and bilinguals. Furthermore, primary caregiver's years of education and children's nonverbal IQ did not appear to influence rule induction in the presence or absence of distractor stimuli. The influence of social–environmental factors, like primary caregiver's years of education, on statistical learning has rarely been considered in prior work. Findings in this study, coupled with previous studies on procedural memory (e.g., Leonard et al., 2015), suggest that some cognitive and environmental factors, such as primary caregiver's years of education (e.g., Eghbalzad et al., 2021), children's nonverbal IQ (e.g., Gebauer & Mackintosh, 2007; Kaufman et al., 2010), and bilingual language experience (e.g., Bulgarelli et al., 2019; Poepsel & Weiss, 2016; Yim & Rudoy, 2013), may have little influence on the development of statistical learning mechanisms. In a post hoc analysis, we removed primary caregiver's years of education and nonverbal IQ scores as covariates and observed the same pattern of results. Despite significant differences in English language ability, primary caregiver's years of education, and nonverbal IQ between monolinguals and bilinguals, language group membership did not predict AGL performance. With and without the use of covariates, monolingual children and bilingual children performed almost identically to each other under light and heavy attention conditions, providing compelling evidence that nonverbal rule induction may be impervious to the influence of social–environmental factors and bilingual language experience.

Summary and Conclusions

Our study is unique in that it considers both the role of attention and bilingualism in statistical learning—factors that have previously been considered separately. Understanding the role of attention and experience in statistical learning is important because it enables us to posit constraints on the process of language acquisition. We observed a significant effect of increased attentional load on performance on a rule-induction task, such that children's performance in the high attention demand condition was significantly worse than performance in the low attention demand condition. This finding suggests that rule induction from visual input is attenuated by the presence of competing visual information.

Future work will need to explore how attentional control may modulate statistical learning performance in the presence of interference across different tasks and modalities. Multiple measures of attention should be explored as a means of identifying the role of domain-general attention mechanisms versus the domain-specific attention mechanisms in statistical learning. In addition, future work will need to examine how different statistical learning paradigms (i.e., AGL, word segmentation, serial reaction time tasks) function under a heavy attention load. Furthermore, the effects of bilingualism as they relate to age of acquisition, current exposure, and proficiency across bilinguals' two languages should be explored when inhibition of irrelevant information is necessary for rule induction on a statistical learning task. As it stands, the current study contributes to the debate on the role of attention in statistical learning and suggests a differential effect of attention on rule induction depending on children's language ability. Coupled with an emerging body of work focused on defining theoretical relationships between attention and statistical learning, as well as language ability and statistical learning, our findings provide a possible pathway to a theoretical model that can consider all three constructs—statistical learning, attention, and language skills—and situate such a framework within a linguistically diverse framework. Such an approach will promote more comprehensive and inclusive models of language acquisition.

Appendix

Symbol Sequences for Grammar 1 and Grammar 2

Grammar 1 Training and Test Sequences

Grammar 2 Training and Test Sequences

Funding Statement

This research was supported by National Institutes of Health Grants R01 DC011750 (awarded to Margarita Kaushanskaya), U54 HD090256 (awarded to Qiang Chang), and F31 DC019025 (awarded to Kimberly Crespo).