Learning of Grammar-Like Sequences by Children With Developmental Language Disorder: An Electrophysiological Examination

Abstract

Purpose:

Developmental language disorder (DLD) is a prevalent neurodevelopmental condition that affects approximately 7% of children in the United States. One notable characteristic of DLD is difficulty in mastering nonadjacent dependencies, such as the third-person singular –s or passive construction be –ed. This study examines behavioral and electrophysiological measures of nonadjacent dependencies learning in children with DLD in the context of a visual grammar-like learning paradigm.

Method:

Fifty-one children (22 DLD, 29 typical language development [TD]) aged 8;0–12;4 (years;months) completed an aXb/cXd grammar learning task, followed by an immediate and 1-week follow-up test. Neurophysiological responses to grammatical and nongrammatical stimuli were measured by recording electroencephalography during the immediate test.

Results:

Children with and without DLD performed similarly and above chance level during the immediate test. Performance on the follow-up test indicated improved accuracy in children with TD. The DLD group, on the other hand, no longer exhibited above chance–level performance, pointing to limited consolidation. Event-related potentials extracted in relation to the grammaticality of the test items highlighted processing differences in responses to grammatical violations between the two groups (TD and DLD). While a P600 was found sensitive to grammatical violations in the TD group, a frontal N400-like activation was found sensitive to violations in the DLD group. These findings are interpreted to suggest familiarity-based associative learning rather than rule-based processing in children with DLD.

Conclusion:

Children with DLD achieved comparable scores on the immediate behavioral test as their TD peers but did not show evidence of rule abstraction and long-term retention.

Developmental language disorder (DLD) is a prevalent neurodevelopmental condition that affects approximately 7% of the children in the United States (Norbury et al., 2016). DLD manifests as significant receptive and expressive language difficulties in the absence of deafness, intellectual disability, brain injury, or other known neurological deficits, and its impact on learning and academic performance persist throughout the lifespan. Children with DLD often encounter considerable barriers across various language domains such as the lexicon, morphology, syntax, discourse (Leonard, 2014), reading (McArthur et al., 2000), and spelling (Joye et al., 2020). Additionally, this disorder frequently co-occurs with other cognitive difficulties in attention (Baker & Cantwell, 1987; K. L. Mueller & Tomblin, 2012) and working memory (Ebert & Kohnert, 2011). A critical area of research in DLD has focused on identifying the underlying impaired learning mechanisms in this disorder to achieve theoretical consensus and optimize interventions.

A key process in language acquisition and other cognitive skills is the ability to detect statistical regularities in the environment, known as statistical learning (SL). SL is an implicit, automatic, and unconscious process that allows learners to extract patterns from continuous sensory input without explicit instruction (Aslin & Newport, 2012). In the context of language, SL enables infants and children to segment words from fluent speech, identify permissible sound combinations, and generalize grammatical structures based on exposure to regularities in phonology, syntax, and morphology.

Research on SL has shown its broad applicability to language development across multiple modalities. For instance, in auditory SL, infants as young as 8 months and adults can segment words from an artificial speech stream based solely on transitional probabilities between syllables and tone sequences (Conway & Christiansen, 2005; Creel et al., 2004; Dawson & Gerken, 2009; Gottselig et al., 2004; Saffran et al., 1996, 1999). In the visual domain, SL supports efficient regularity identification, facilitates future predictions, and therefore contributes to faster and more efficient decoding, reading fluency, and reading comprehension (Arciuli & Simpson, 2012; Aslin & Newport, 2012; Fiser & Aslin, 2002; Pavlidou & Bogaerts, 2019; Reber, 1967). SL also operates in the visuomotor domain, to support learning of complex motor tasks through repeated exposures (Cleeremans & McClelland, 1991; Lammertink et al., 2020). Given its essential role in language acquisition, numerous studies have investigated SL performance in individuals with DLD.

Deficits in recognizing statistical patterns in individuals with DLD have been found in a variety of linguistic and nonlinguistic contexts, such as statistical word learning tasks (Ahufinger et al., 2022; Evans et al., 2009; Hsu & Bishop, 2014; Mayor-Dubois et al., 2014), serial reaction time tasks (Gabriel et al., 2013; Hsu & Bishop, 2014; Lum et al., 2010, 2012; Tomblin et al., 2007), and artificial grammar learning tasks (Gillis et al., 2022). However, some studies found no difference in SL between participants with and without DLD (e.g., serial reaction time tasks: Gabriel et al., 2011; Hedenius et al., 2011; visual triplet tasks: Lammertink et al., 2020; Noonan, 2018; and word segmentation task: Noonan, 2018). Further investigation is needed to determine specific factors influencing SL performance in DLD. Sequence learning, in particular, has received growing attention. Hsu and Bishop (2014) identified sequence-specific deficits in children with DLD, by comparing their performance on two motor procedural learning tasks and one verbal sequence learning task. Additionally, SL studies on nonadjacent dependencies in adults with and without DLD using auditory (Grunow et al., 2006; Hsu et al., 2014) and visual stimuli (Aguilar & Plante, 2014) highlighted the importance of investigating learning of order and co-occurrence dependencies.

Nonadjacent dependencies in language refer to grammatical or semantic relationships between two elements that are separated by one or more intervening elements within a sentence. The two elements must correspond or relate in a certain way. For example, in the sentence “The dogs in the yard are barking,” the subject “dogs” and the verb “are” agree in number, despite being separated by the prepositional phrase “in the yard.” Such structures are critical for understanding syntax and hierarchical rules in natural languages, as they extend beyond simple linear order. Successful processing of nonadjacent dependencies requires cognitive ability and working memory capacity to track dependencies across distance. Studying nonadjacent dependencies in the context of DLD is important for understanding their syntactic learning patterns.

The aXb/cXd grammar-like SL task has emerged as a useful tool for investigating nonadjacent dependencies. In this paradigm, participants are exposed to structured input sequences composed of triplets in which the first and third elements are consistently paired (e.g., symbol “a” always comes first and is paired with symbol “b”), while the middle element “X” varies randomly. Successful learning is reflected by discriminating new grammatical strings (i.e., aXb, cXd with new “X”s) from nongrammatical ones, through the detection of grammatical violations (i.e., co-occurrence violations: aXd, cXb; linear-order violations: bXa, dXc). This structure simulates commonly used patterns in English grammar (e.g., the passive construction “be –ed” and the third-person singular “He/She –s”) and therefore provides a window into aspects of language learning patterns and deficits. The use of artificial language stimuli allows better control over prior learning and exposure conditions. In this paradigm, all participants are assigned to the same learning environment, and the amount of exposure to grammatical stimuli is carefully balanced. The passive listening or observation eliminates the impact of motor skills that may be affected in DLD on performance (Sack et al., 2022; Tseng & Hsu, 2023). Finally, this design allows for testing the generalization of rules to new exemplars, an important factor of SL compared to surface familiarization (Daltrozzo & Conway, 2014).

Aguilar and Plante (2014) studied nonadjacent SL with a visual aXb/cXd task in adults with and without language learning disability (LLD) and found no evidence of a deficit in adults with this disorder. These findings were in contrast with previous findings of difficulties with auditory nonadjacent SL in adults with a language learning disorder (Grunow et al., 2006). There are several possible factors that could explain why individuals with language disorders have not always shown deficits in SL tasks. First, even though SL tasks are designed to examine implicit learning, explicit processing such as memorization may not be entirely excluded, which might play a supplementary or distractive role (Daltrozzo & Conway, 2014). It is suggested that individuals with DLD compensate for their impaired implicit learning by relying on relatively intact declarative memory to support learning (Evans et al., 2022). It is therefore possible that in Aguilar and Plante (2014), adults with LLD applied explicit strategies to cope with their learning challenges, which may have masked their deficits in SL in this task.

Second, behavioral assessments of SL often rely on explicit responses and might not fully capture subtle nuances in cognitive processing. Behavioral responses can be subject to various confounding factors (Bell & Cuevas, 2012), such as individual differences in decision-making preferences influenced by anxiety levels and self-esteem (Burnley et al., 2024). Electroencephalography (EEG) is a valuable tool for studying implicit processing by measuring event-related potentials (ERPs). EEG has been used as an effective measurement in a variety of SL tasks (Bahlmann et al., 2006; Lelekov et al., 2000; Schröger et al., 2007; Silva et al., 2017, 2018).

ERP Components Associated With the Detection of Grammatical Violations

Derived from EEG data, ERPs are synchronized neural activities in the brain observed in relation to specific events and reflect distinct cognitive processing. A key advantage of ERPs is that they can be detected without conscious awareness or overt behavioral responses, making them particularly useful for investigating implicit learning mechanisms such as SL (Luck, 2005). ERPs can provide more detailed information about stimulus processing during SL, even when behavioral indicators are absent or minimal, and offer evidence on the shared mechanisms between SL and language processing. The P300 is an ERP component that is particularly relevant to the study of implicit learning because it indexes the ongoing, often unconscious, integration of new information into an evolving internal model of the environment (Donchin & Coles, 1988).

The P300 is a positive-going ERP component that peaks at least 300 ms after stimulus onset, is maximal over parietal scalp regions, and is elicited by task-relevant rare events in the context of frequent events (Picton, 1992; Sutton et al., 1965). The amplitude of the P300 is influenced by stimulus probability, task relevance, and expectancy, with larger amplitudes elicited by rare, unexpected, or meaningful stimuli that require updating of internal representations (Donchin, 1981; Donchin & Coles, 1988; Johnson, 1986; Polich, 2007). The latency of the P300 is associated with the time required for stimulus evaluation, increasing when stimulus categorization is more demanding (Kutas et al., 1977). Within the context of language processing, a P300-like positivity with a latency of approximately 600 ms has been shown to be sensitive to syntactic violations and is commonly referred to as the P600 (Hagoort et al., 1993).

The P600 amplitude has been reported to be larger in response to syntactic violations (Hagoort et al., 1993; Osterhout & Holcomb, 1992); complex (Friederici et al., 2002; Kaan et al., 2000), less preferred syntactic structures (Gouvea et al., 2010; Itzhak et al., 2010; Osterhout et al., 1994); and long-distance structural dependencies (Felser et al., 2003; Phillips et al., 2005) that require repair or reanalysis. The P600 is also sensitive to grammatical violations in artificial languages such as artificial grammar learning tasks (Silva et al., 2017, 2018) and sequential learning tasks with grammar-like underlying structures (Christiansen et al., 2012). Specifically, in Silva et al.'s (2017) Artificial Grammar Learning study, the P600 was not found to be sensitive to surface familiarity, suggesting that it reflects a response to structural irregularities.

Within the context of our study, a P600/P300 elicited by syntactic violations following exposure to exemplars with an aXb cXd structure would reflect learning of the underlying rules. A stronger P600 response to grammatical violations would indicate detection of deviations from the learned structure, reflecting the updating of internal grammatical representations and providing evidence for the acquisition of abstract structural knowledge.

The N400 is an ERP component that peaks around 300–500 ms after stimulus onset and is primarily observed over central regions of the scalp, reflecting the processing of meaning. Its amplitude has been found to be modulated by the difficulty with which meaning can be integrated into a given context (Hagoort, 2007; Kutas & Federmeier, 2011). The N400 was reported to be larger when a word is semantically unexpected or incongruent with the context (see Kutas & Federmeier, 2011). N400 effects were found across input modalities, including spoken and written words, American Sign Language signs, and pseudowords (see Kutas & Van Petten, 1990), but not in other structured domains such as music (Besson & Macar, 1987). The spatial distribution of this negativity varies across reports, with a frontal distribution reported when the N400 is elicited by pictures (Ganis et al., 1996), and occipital maximum when the negativity is triggered by mismatched faces (Olivares et al., 1999). Evidence exists to suggest that the N400 may reflect broader mechanisms of expectancy violation and violations in pattern learning in artificial grammar paradigms (J. L. Mueller et al., 2009; Tabullo et al., 2013). Therefore, our study also targeted the N400 as a potential measure of SL among children with and without DLD.

The Present Study

The present study aimed to examine the learning process of nonadjacent dependency regularities in an aXb/cXd paradigm by children with DLD. We hypothesized that unlike adults with language learning disorders, children with DLD do not employ effective compensatory explicit strategies (Daltrozzo & Conway, 2014) and therefore will exhibit SL difficulties. Electrophysiological measurements (ERPs) were employed to capture neural activity associated with explicit learning and SL. A 1-week follow-up behavioral test was included to assess long-term retention and consolidation, as knowledge acquired through SL is typically robust and rapidly processed compared to information learned explicitly. Two research questions were addressed in the present study:

1. Do (and in what way) children with DLD differ from their age-matched typically developing peers in the learning outcomes and retention of grammatical regularities in aXb cXd learning?

2. What ERP patterns (e.g., P600, N400) are associated with the processing of grammatical violation processing following aXb cXd learning in children with and without DLD?

Method

Participants

Children aged 8;0–12;4 (years;months), from the greater Boston area, were recruited to participate in the study. All participants had normal or corrected-to-normal vision, and they had no history of neurological deficits or head injuries based on the caregiver's report. English was the predominant language for all participants. We verified that participants had no consistent exposure in the home or school environment to the languages from which the a, b, c, and d symbols used in the experimental paradigm were derived to avoid inadvertent engagement of semantic processing of these languages. Because of the great variability of X stimuli that were drawn from multiple sources and languages and their sporadic presentations throughout the experiment, there was no concern that participants will rely on familiarity with symbols to engage with the task. Fifty-one children completed the study, including 22 participants with DLD and 29 age-matched controls with typical language development (TD). The DLD group had an M_age of 120.55 months (SD = 14.62, minimum: 99 months, maximum: 148 months). The TD group had an M_age of 118.97 months (SD = 15.33, minimum: 96 months, maximum: 143 months).

Children in the DLD group met the following criteria: (a) an Identification Core score on the Test of Integrated Language and Literacy Skills (TILLS; Nelson et al., 2016) of less than 34 for ages 8–11 years or less than 42 for ages 12–18 years; or (b) the scaled score of at least two subtests of TILLS is below the average range. All participants completed a nonverbal IQ screening through the Matrices subtest of the Kaufman Brief Intelligence Test–Second Edition (KBIT-2; Kaufman & Kaufman, 2004) and had a nonverbal intelligence score above the range of intellectual disability (standard score > 70). Participants completed the Number Repetition and Familiar Sequences subtests of the Clinical Evaluation of Language Fundamentals–Fourth Edition (CELF-4; Semel, Wiig, & Secord, 2003) as supplemental measure of verbal working memory. The means and standard deviations of test scores for each group are presented in Table 1. As expected, the TD group showed higher scores than the DLD group on measures of language ability (TILLS Identification Core), t(49) = 9.04, p < .001, Cohen's d = 2.55, and verbal working memory (CELF-4 Working Memory Index), t(49) = 3.90, p < .001, Cohen's d = 1.10. The TD group also showed higher scores on nonverbal reasoning (KBIT-2 Matrices) than the DLD group, t(49) = 3.66, p < .001, Cohen's d = 1.03, which is not surprising given consistent reports of nonverbal IQ scores that are within the normal range but lower than those of peers among children with DLD (Earle et al., 2017; Gallinat & Spaulding, 2014; Newton et al., 2010; Saar et al., 2018).

Table 1.

Participant data by group.

Variable	DLD (n = 22) M (SD)	TD (n = 29) M (SD)
Age in months	120.55 (14.62)	118.97 (15.33)
KBIT-2 matrices score	105.14 (15.16)	119.90 (13.56)
TILLS Identification Core score	75.27 (14.13)	106.72 (10.75)
CELF-4 Working Memory Standard Score	93.73 (16.51)	109.24 (11.90)
Gender	9 females, 13 males	12 females, 16 males, 1 nonbinary

Note. DLD = developmental language disorder; TD = typical development; KBIT-2 = Kaufman Brief Intelligence Test–Second Edition; TILLS = Test of Integrated Language and Literacy Skills; CELF-4 = Clinical Evaluation of Language Fundamentals–Fourth Edition.

Procedure

This study was approved by Mass General Brigham Institutional Review Board (Protocol #: 2020P001328). Parent consent and participant assent were obtained before data collection, and participants were compensated for their participation. Participants were enrolled in a larger study on the relationship between implicit and explicit learning in children with DLD, which included three sessions lasting 2 hr each. Standardized assessments were administered in one to two sessions prior to completing the experiment tasks. The experimental task of the reported study was completed in a single session.

Participants were fitted with a 32-channel EEG HydroCel net by Electrical Geodesics, Inc., and seated in a quiet room, at a comfortable distance (60 cm) from a computer monitor adjusted to align the center of the screen with the participant's eye level. We used programmable experiment generation software (E-Prime 2.0, Psychological Software Tools) to control the presentation of stimuli. Each participant completed an aXb/cXd task, lasting about 15 min, while their EEG data were recorded from the scalp. One week after the in-person session, a remote follow-up test built in RedCap was sent via e-mail for participants to complete on their home devices. All children and caregivers were provided with specific instructions and practiced the procedure of the follow-up test during their in-person session to ensure a similar level of testing conditions.

Task

The aXb/cXd task consisted of two phases, the learning phase and the test phase. The learning phase took approximately 10 min. Participants were presented with 160 aXb cXd strings on a computer monitor, with each string displayed for 1,000 ms. A fixation with a cross symbol at the center of the screen was displayed for 1,000 ms between strings. Participants were instructed to pay close attention to the visual stimuli on the screen, without prompting them toward active memorization. Two self-paced short breaks were provided during the learning phase.

Immediately following the learning phase, participants completed a test phase in which they were required to indicate whether new strings were consistent with the rules of the language they were previously exposed to. Each string was presented on the screen for 1,000 ms, followed by a red question mark added to cue participants to indicate whether the new string was “correct” or “incorrect.” Infinite time was given until participants made the judgment by pressing one button on a response box. A fixation with a cross symbol at the center of the screen was displayed for 1,000 ms between strings. Responses were collected through a Chronos response box, with no feedback given during this phase. The test consisted of 96 strings and took approximately 5 min to complete. One week after the in-person completion of the task, participants were required to complete a follow-up test remotely, consisting of 20 strings they were asked to judge as correct or incorrect.

Stimuli

Visual stimuli were designed based on the structure proposed by Aguilar and Plante (2014; see Figure 1). All grammatical stimuli presented in the experiment reflected an “aXb” and “cXd” grammar in which a–b and c–d elements always occurred as a pair, and the third element (X) occurred in the middle position. The “X” symbols came from a pool of 64 special characters. Forty “X” symbols were presented four times each during the learning phase, twice in the “aXb” sequence and twice in the “cXd” sequence, for a total of 160 sequences. The nongrammatical stimuli used in the test phase were composed of two types of violation, linear order violation (i.e., bXa, dXc) and co-occurrence violation (i.e., aXd, cXb). During the testing phase, 24 new “X” symbols were used four times each, twice in the grammatical (G) condition, once in the nongrammatical condition with linear order violation (NGO), and once in the nongrammatical condition with co-occurrence violation (NGC), composing a total of 96 test strings. The remote follow-up test consisted of 20 sequences balanced across stimulus types.

Figure 1.

Examples of aXb and cXd strings.

Data Analysis

Response Bias

An overall bias toward either “correct” or “incorrect” judgment responses can artificially inflate or reduce overall accuracy rates, potentially leading to misleading group differences in behavioral results. However, this evaluation should be applied with caution because variability in detecting different types of violations may give the wrong impression of a response bias (e.g., successfully acquiring the order rule but not the co-occurrence rule could result in having 75% of stimuli judged as “correct”). Therefore, in the context of our paradigm, response bias should meet two criteria: (a) a general response bias and (b) response bias that is not unique to only one of the two grammatical violations.

We assessed the general response bias using Macmillan and Creelman's (2004) equation: C = −[z(Hit) + z(FA)]/2, where the sum of the z scores of hits (correct endorsements of grammatical strings) and false alarms (incorrect endorsements of nongrammatical strings) is calculated. A negative C value indicates a bias toward responding “correct” (i.e., grammatical) more than “incorrect” (i.e., nongrammatical), while a positive C value indicates the opposite bias. Participants with C values more than 1.5 SDs below or above the mean were considered to have a bias. In addition, if a participant's false alarm rate in detecting order violations and co-occurrence violations differed by no more than 0.25, they were considered to have an overall response bias not attributable to learning to detect only one of the violations.

Behavioral Data Analysis

Statistical analyses for behavioral and ERP data were conducted using R (Version 4.4.1, RStudio 2024.04.2). First, discriminability (d') was calculated by comparing hit rates (proportion of grammatical strings correctly endorsed as grammatical) and false alarm rates (proportion of nongrammatical strings judged incorrectly as grammatical). Learning is indicated as the size of the difference between hits and false alarms, with hits outweighing false alarms reflect better learning. d' values of each group were compared to zero with a one-sample t test to evaluate whether learning took place in the immediate and follow-up tests, respectively. An independent t test was used to compare the group difference in d' values. When the assumption of normality was violated (p < .05), Wilcoxon tests and Welch's t tests were employed as nonparametric alternatives.

To further examine the effects of stimulus type and time of test, a 2 × 2 × 3 mixed-design analysis of variance (ANOVA) was completed with the group (TD, DLD) as the between-subjects variable, stimulus type (G, NGC, NGO) and time of test (immediate, follow-up) as the within-subject variables, and test accuracy as the dependent variable. Greenhouse–Geisser corrections were applied to degrees of freedom when Mauchly's test of sphericity indicated a violation of the sphericity assumption (p < .05). Post hoc t tests with Bonferroni corrections were conducted to further analyze significant effects observed in the ANOVAs.

EEG Data Acquisition and Signal Processing

The 32-channel GES 400 System by Electrical Geodesics, Inc. was used to record the EEG data, using 32-channel HydroCel Geodesic sensor nets composed of Ag/AgCl electrodes attached to an elastic net following the International 10–20 system. EEG was continuously recorded at a sampling rate of 1000 Hz with the vertex as the reference electrode, with built-in anti-aliasing filters (high-pass: 0.1 Hz, and low-pass: 500 Hz) to ensure that no frequency content above half the sampling rate was aliased into the passband. Impedances were kept below 50 kΩ as per the manufacturer's recommendations (Ferree et al., 2001). As reported in the Task section above, the stimulus onset asynchrony varies due to the infinite time given for response. The intertrial interval is 1,000 ms in this task. No temporal jitter was introduced in the design. EEG was time-locked to the onset of each string, with a baseline correction window from −200 to 0 ms relative to stimulus onset. Custom MATLAB (The MathWorks, Inc.) scripts operating in conjunction with the open-source EEGLAB toolbox (Delorme & Makeig, 2004; http://sccn.ucsd.edu/eeglab) were used to perform the EEG data processing and analysis.

A bandpass filter of 0.1–40 Hz was used to filter the continuous data for ERP analyses. The processed data were segmented into 1,200-ms epochs extending 200 ms before and 1,000 ms after the stimulus onset. Each trial was visually inspected for movement artifacts and other signal noise, and trials with artifacts were manually removed. Baseline correction was performed on the averaged data, based on the signal in the 200-ms preceding the onsets of the visual stimuli (i.e., −200 to 0 ms). Data were rereferenced to the average reference (Winkler et al., 2015). An adaptive mixture Independent Component Analysis was applied separately to single-subject data sets (Delorme et al., 2012) to detect and correct for eye movement and blinks. The data were then sorted into three categories based on the stimulus types (G, NGO, and NGC) for each group (TD and DLD). In the TD group, 7.5% of the G trials, 8.5% of the NGC trials, and 7.6% of the NGO trials were rejected. In the DLD group, 9.2% of the G trials, 11.5% of the NGC trials, and 10.4% of the NGO trials were rejected. The ERP analysis was performed on trials from the testing phase only, and all trials were included in the analysis regardless of participants' response accuracy.

EEG Data Analysis

A visual inspection of the ERP waveforms as well as the topographic maps (see Figure 2 and Figure A1) indicated maximal sensitivity to grammaticality in the centro-parietal site (Pz), consistent with the P600 and at the midline frontal site (Fz) consistent with a frontal N400. Data analysis of these two well-established ERP components of interest (P600, N400) was conducted on the midline electrodes (Fz, FCz, Cz, and Pz) and two time windows of interest (300–500 ms for the N400, and 500–900 ms for the P600). To reduce the temporal dimensionality of the data set and to disentangle ERP components that overlapped in time, temporal principal component analyses (TPCA) with a Promax rotation (Dien, 2010) were conducted (Arbel et al., 2017; Arbel & Wu, 2016) on the midline electrodes (Fz, FCz, Cz, and Pz). The TPCA used the covariance between time points and resulted in a set of 11 temporal factors accounting for the maximal amount of the total variance (78.61%). The peak of Temporal Factor 1 (TF1) was consistent with the peak of P600 observed in the grand average waveforms at Pz (see Figure 4). Temporal Factor 2 (TF2) with a maximal peak around 400 ms was consistent with the peak of frontal N400 observed in the grand average waveforms at Fz (see Figure 5). Temporal factor scores were extracted for the statistical analyses.

Figure 2.

Left: Event-related potential (ERP) waveforms of midline electrodes (Fz, FCz, Cz, and Pz) in the typical language development (TD) group; central top: topographic maps of each group (TD, developmental language disorder [DLD]) and stimulus type (grammatical [G], co-occurrence violation [NGC], linear order violation [NGO]) in the time window of FN400 (300–500 ms). Central bottom: topographic maps of each group (TD, DLD) and stimulus type (G, NGC, NGO) in the time window of P600 (500–900 ms). Right: ERP waveforms of midline electrodes in the DLD group.

Figure 4.

Top left: Event-related potential (ERP) magnitudes across group (typical language development [TD] vs. developmental language disorder [DLD]), stimulus type (grammatical [G] vs. co-occurrence violation [NGC] vs. linear order violation [NGO]), and electrode (Fz vs. FCz vs. Cz vs. Pz) from Temporal Factor 1 (TF1). Error bars represent the standard error of the mean. Bottom left: Principal Component Analysis plot with a set of 11 temporal factors accounting for the total variance at midline electrodes, a bold line for TF1. Bottom center: ERP waveforms of differences (NGC − G; NGO − G in both groups) in the time window of P600. Right: ERP waveforms of P600 across group (DLD, TD) and stimulus type (G vs. NGC vs. NGO) at Pz electrode.

Figure 5.

Top left: Event-related potential (ERP) magnitudes across group (typical language development [TD] vs. developmental language disorder [DLD]), stimulus type (grammatical [G] vs. co-occurrence violation [NGC] vs. linear order violation [NGO]), and electrode (Fz vs. FCz vs. Cz vs. Pz) from Temporal Factor 2 (TF2). Error bars represent the standard error of the mean. Bottom left: Principal Component Analysis plot with a set of 11 temporal factors accounting for the total variance at midline electrodes, a bold line for TF2. Bottom center: ERP waveforms of differences (NGC − G; NGO − G in both groups) in the time window of FN400. Right: ERP waveforms of FN400 across group (DLD, TD) and stimulus type (G vs. NGC vs. NGO) at Fz electrode.

We conducted mixed-design ANOVAs to examine the effects and interactions of electrode (Fz, FCz, Cz, and Pz), group (DLD vs. TD), and stimulus type (G, NGC, NGO) on the corresponding temporal factor scores, to depict the electrodes of maximal activation for each ERP component of interest (P600, N400). We then evaluated the effect and interactions of group (DLD vs. TD) and stimulus type (G, NGC, NGO) associated with each ERP component of interest. Greenhouse–Geisser corrections were applied when assumptions of sphericity were violated. Post hoc paired t tests were conducted when effects or interactions were significant (p < .05).

Results

Behavioral Results

Response Bias

The average criterion for the sample was C = −0.34 (SD = 0.52), indicating an overall bias toward responding “correct” (i.e., grammatical). Three participants with criterion values more than 1.5 SDs below or above the mean (i.e., C < −1.12 or C > 0.43) were excluded from further analyses. One participant from the TD group had a bias toward “wrong” (i.e., C = 0.72), and two from the DLD group had a bias toward “correct” (i.e., C = −1.16, C = −2.57). Data from the remaining 48 participants (20 DLD, 28 TD) were included in the following analyses.

Learning Effect According to Overall Performance

Immediate test. We first evaluated the d' (d-prime) scores for each group. A one-sample Wilcoxon test indicated that the d' of the DLD group (Mdn = 0.35) was significantly greater than zero, V = 160, p < .01. A t test indicated that the d' of the TD group (M = 0.94, SD = 1.09) was also significantly greater than zero, t(27) = 4.79, p < .001. A Welch's two-sample t test revealed no significant difference between the groups, t(45.99) = −1.91, p = .06. The effect size, Cohen's d = −0.53, indicated a moderate effect. These results suggest that hits were significantly more prevalent than false alarms in both groups, with no significant difference between the groups. In summary, both DLD and TD groups demonstrated learning effects, as evidenced by their d' scores with no significant differences between the groups in these measures.

Follow-up test. Wilcoxon tests indicated that the d' was not significantly greater than zero for the DLD group (Mdn = 0.13), V = 91.5, p = .11, but significantly greater than zero for the TD group (Mdn = 1.80), V = 280.5, p < .001. A Welch's two-sample t test revealed a significant difference between the groups, t(36.28) = −3.97, p < .001, suggesting that the TD group reached a greater d' than the DLD group on the follow-up test. The effect size, Cohen's d = −1.02, indicated a large effect.

Effect of Group, Violation Type, and Time of Test on Task Performance

A 2 × 2 × 3 mixed-design ANOVA was conducted to examine the effect of group (DLD and TD), stimulus type (G, NGC, NGO), and time of test (immediate, follow-up) on test accuracy. The results are presented in Figure 3.

Figure 3.

Mean accuracy on grammatical, linear order violation, co-occurrence violation test strings by participant groups and time of test. G = grammatical; NGC = co-occurrence violation; NGO = linear order violation; DLD = developmental language disorder; TD = typical language development.

Group. The analysis revealed a significant main effect for group, F(1, 46) = 8.34, p < .01, η²_G = .05, indicating that overall accuracy regardless of test timing (immediate and follow-up) was greater in the TD group. While there was no main effect of time, F(1, 46) = 2.75, p = .10, η²_G = .004, an interaction between group and time was discovered, F(1, 46) = 6.09, p < .05, η²_G = .008. The post hoc paired t tests indicated that the DLD group showed no significant difference between immediate (M = 0.57 ± 0.09) and follow-up accuracy (M = 0.54 ± 0.10), t(19) = 1.39, p = .18, Cohen's d = 0.28, while the TD group experienced a significant increase in accuracy (immediate: M = 0.65 ± 0.16, follow-up: M = 0.71 ± 0.22), t(27) = −2.37, p < .05, Cohen's d = −0.33, suggesting a small effect. Taken together, the group effect together with the Group × Time interaction was driven by an increase in accuracy in the TD group from the immediate to the follow-up test, while the DLD group showed no significant change.

Stimulus type. The ANOVA also revealed a significant main effect for stimulus type after correction, F(1.28, 58.05) = 14.76, p < .001 η²_G = .13. Post hoc tests showed that the accuracy on G strings across groups and time was significantly greater than that on NGC strings, t(95) = 5.09, p < .001. The effect size, Cohen's d = 0.82, suggested a large effect size. The accuracy on NGO strings was also found to be significantly greater than that of NGC strings, t(95) = 5.02, p < .001, Cohen's d = 0.78 (moderate-to-large effect). There were no accuracy differences between G and NGO strings, t(95) = −0.12, p = .90, Cohen's d = −0.01 (negligible effect). These results suggest that participants were better at detecting grammatical strings and order violations than strings with co-occurrence violations.

An interaction between stimulus type and time was significant, F(1.71, 78.67) = 7.38, p < .01, η²_G = .03. Post hoc tests revealed that the interaction is driven by a significant increase in accuracy on NGC strings only from a below-chance level (M = 0.38 ± 0.20) in the immediate test to an above-chance level (M = 0.55 ± 0.35) in the follow-up test, t(47) = −3.41, p < .01, Cohen's d = −0.59. No significant differences between the immediate and follow-up test were found for G strings, t(47) = 0.52, p = .61, Cohen's d = 0.07, or NGO strings, t(47) = 0.86, p = .40, Cohen's d = 0.11.

Group × Stimulus Type interaction, F(1.28, 58.05) = 1.74, p = .19, η²_G = .02, and Group × Stimulus Type × Time interaction, F(1.71, 78.67) = 0.56, p = .55, η²_G = .002, were not significant.

In summary, the results revealed that the TD group performed better than the DLD group on the follow-up test, despite no group differences on the immediate test. The TD group also exhibited a significant increase in accuracy from the immediate to the follow-up test, while the DLD group exhibited no change. Finally, the accuracy for NGC strings increased from below chance in the immediate test to above chance in the follow-up test, regardless of group.

Rule Extraction During the Follow-Up Test

Examination of individual performance on the follow-up test revealed patterns consistent with two subgroups of learners, one that learned the order rule, and the other that showed consolidation of both rules at follow-up.

Partial rule extractors. Six participants (five TD, one DLD) correctly endorsed almost all G strings and rejected all NGO strings but incorrectly endorsed all NGC strings. This suggests that they extracted only the order rule and used it to guide their judgment in the follow-up test consistently (see Table 2).

Table 2.

Performance of partial rule extractors.

ID	Group	Immediate test accuracy			Follow-up test accuracy
		G	NGC	NGO	G	NGC	NGO
11	DLD	1	0.04	0.96	1	0	1
12	TD	0.63	0.46	0.63	1	0	1
15	TD	0.79	0.17	0.83	1	0	1
18	TD	0.96	0.21	0.96	0.9	0	1
35	TD	0.98	0	1	1	0	1
48	TD	0.83	0.08	0.92	1	0	1

Note. G = grammatical; NGC = co-occurrence violation; NGO = linear order violation; DLD = developmental language disorder; TD = typical development.

Complete rule extractors. Nine participants from the TD group endorsed almost all G strings and correctly rejected almost all NGC and NGO strings (see Table 3). This suggests that they extracted both order and co-occurrence rules and used them to guide their judgment on the follow-up test consistently. Interestingly, although these learners achieved near-perfect accuracy for the NGC type in the follow-up test, none demonstrated clear evidence of acquiring this rule during the immediate test (i.e., accuracy < 0.55).

Table 3.

Performance of complete rule extractors.

ID	Group	Immediate test accuracy			Follow-up test accuracy
		G	NGC	NGO	G	NGC	NGO
7	TD	0.88	0.54	0.96	1	1	1
13	TD	0.96	0.50	0.96	1	1	1
14	TD	0.88	0.46	0.83	1	1	1
19	TD	0.90	0.42	1	1	1	1
20	TD	0.77	0.13	0.92	1	1	1
30	TD	0.85	0.54	0.96	0.9	1	1
31	TD	0.98	0.42	0.92	1	0.8	1
39	TD	0.90	0.50	1	1	1	1
40	TD	0.96	0.42	1	0.9	1	1

Note. G = grammatical; NGC = co-occurrence violation; NGO = linear order violation; TD = typical development.

Individual Differences in SL Outcomes

To explore potential predictors of individual differences in SL outcomes, we conducted a multiple linear regression analysis. The dependent variable was participants' SL outcomes, represented by d' scores. The independent variables included age (in months), language skills assessed by the TILLS standard scores, and working memory capacity measured by the Core Language Score from the CELF-4 (Semel, Wiig, & Secord, 2003).

The regression model was statistically significant, F(3, 44) = 3.81, p < .05, and accounted for approximately 20.6% of the variance in d' scores (R² = .21, adjusted R² = .15). Among the evaluated predictors, the TILLS standard score was a significant contributor, β = .02, SE = 0.01, t = 2.72, p < .01, indicating that stronger language skills were associated with enhanced SL outcomes. Age also exhibited a trend toward significance, β = .02, SE = 0.01, t = 1.91, p = .06, indicating that older participants might tend to perform better on the immediate test. Working memory was not found to be a significant predictor of SL outcomes (β = −0.02, SE = 0.013, t = −1.67, p = .10).

We then conducted a second regression analysis using d' scores from the follow-up test as the dependent variable, with the same set of predictors. The model was also statistically significant, F(3, 44) = 4.18, p < .05, and accounted for approximately 22.2% of the variance in d' scores (R² = .22, adjusted R² = .17). In this model, TILLS standard scores maintained as a significant predictor, β = .04, SE = 0.02, t = 2.70, p < .01, indicating that participants with stronger language skills tended to show better long-term learning effects in the follow-up test. Age and working memory were not significant predictors, age: β = .02, SE = 0.02, t = 1.27, p = .21, working memory: β = −0.01, SE = 0.02, t = −0.46, p = .65.

ERP Results

P600

A mixed-method ANOVA was conducted to evaluate the effects of group (DLD, TD), stimulus type (G vs. NGC, NGO), and electrodes (Fz, FCz, Cz, and Pz) on the magnitude (factor scores) of the P600 component (TF1). The result indicated a significant electrode effect after a Greenhouse–Geisser correction, F(1.68, 77.36) = 21.24, p < .001, η²_G = .12. Post hoc t tests confirmed that the positive activation within the 500–900 time window was maximal at the Pz electrode (ps < .001), consistent with the targeted P600 component.

A one-way ANOVA at Pz indicated a significant effect of stimulus type on the P600 amplitude, F(2, 54) = 3.17, p < .05, η²_G = .04. Post hoc paired t tests revealed that the magnitude of the P600 was larger for NGO stimuli than for G stimuli, t(27) = 2.88, p < .01, Cohen's d = 0.50. There was no difference in the P600 amplitude between NGC and G stimuli, t(27) = 0.90, p = .38, Cohen's d = 0.15. A stimulus type effect was not found in the DLD group, F(2, 38) = 0.56, p = .58, η²_G = .008. These results suggested a sensitivity of the P600 to order violations in the TD group. The results are presented in Figure 4.

Frontal N400

A mixed-method ANOVA was conducted to evaluate the effects of group (DLD, TD), stimulus type (G vs. NGC, NGO), and electrode (Fz, FCz, Cz, and Pz) on the magnitude (factor scores) of the frontal F400 (TF2). The result indicated a significant electrode effect after correction, F(1.63, 74.99) = 45.95, p < .001, η²_G = .28. Post hoc t tests confirmed that the negative activation within the 300–500 time window was maximal at the Fz electrode (ps < .001), which is consistent with a frontal N400 observed in the grand average data of the DLD group in our sample.

A one-way ANOVA at Fz indicated a significant effect for stimulus type in the DLD group, F(2, 38) = 3.27, p < .05, η²_G = .06. Post hoc paired t tests revealed that the N400 amplitude of both violations (NGC and NGO) in the DLD group was significantly more negative than for G stimuli, NGC-G: t(19) = −2.40, p < .05, Cohen's d = −0.64, NGO-G: t(19) = −2.36, p < .05, Cohen's d = −0.38. No stimulus type effect at Fz was found in the TD group, F(2, 54) = 0.17, p = .84, η²_G = .002. These results suggested that a frontal N400-like negativity in response to grammatical violations was detected only in the DLD group. The results are presented in Figure 5.

Summary of Results

The behavioral results of the immediate test did not support our initial hypothesis that children with DLD would exhibit poorer SL when compared with their peers. Instead, the DLD and TD groups demonstrated a learning effect in the immediate test, with no significant difference between the groups. However, significant group differences emerged during the follow-up test conducted 1 week after the task. The TD group maintained a learning effect and experienced an increase in accuracy, while for the DLD group, the d' score, which was above chance on the immediate test, was no longer different from chance at follow-up. Exploratory regression analyses suggested language skills measured by TILLS as a predictor of both immediate performance and long-term consolidation, while age as a marginal predictor of immediate performance only.

The ERP results revealed that the TD group exhibited a P600 in response to order violations, which was missing in the DLD group. A frontal N400-like response was elicited by both co-occurrence and linear order violations in the DLD group, but not in the TD group.

Discussion

In this study, we utilized the aXb/cXd paradigm to evaluate SL in children with and without DLD. Consistent with Aguilar and Plante's (2014) report on adults with and without LLD, we found no accuracy group differences in the immediate test across the stimulus types. These findings could not be attributed to age, or gender, as there were no significant differences in these variables between the DLD and TD groups.

The results provide evidence of learning in school-age children with and without DLD. Participants showed evidence of applying underlying rules of order they were exposed to during the learning phase to differentiate between grammatical and nongrammatical novel exemplars in the immediate test. This effect was achieved after a rapid exposure phase with no direct learning instructions given. Children with DLD in our study achieved learning outcomes (accuracy: M = 0.57) that are comparable to those reported in adults with LLDs in Aguilar and Plante's (2014) study (accuracy: M = 0.545). It should be noted that in Aguilar and Plante, adults were exposed to 40 unique exemplars, each presented twice (total 80 presentations) during the learning phase, and in our study, children were exposed to the 40 unique exemplars four times (a total of 160 presentation). Because we have not manipulated exposure levels, it is yet to be determined if accuracy in children with DLD is comparable to that of adults with language learning disorders when exposure is the same.

Our results of no differences in d' or accuracy on the immediate test between children with and without DLD are in contrast with Gillis et al.'s (2022) report of inferior performance by children with DLD when compared with peers on a visual artificial grammar learning task. The artificial grammar learning task followed a more complex underlying grammatical structure that did not provide room for explicit strategies to advance learning. Children with DLD had to rely primarily on their implicit learning mechanisms and may have failed to engage effectively explicit learning strategies. The aXb/cXd task, on the other hand, offers a simpler, single-layer underlying grammar, allowing the engagement of a limited explicit learning system to facilitate performance, at least during the initial stages of learning. These disparities suggest that complexity could be a critical factor impacting the ability to use the already limited explicit learning mechanisms to support SL in children with DLD.

While children with and without DLD showed evidence of correctly accepting grammatical strings and identifying linear violations, they did not exhibit evidence of identifying co-occurrence violations in the immediate test. The variations observed in detecting different violation types align with previous reports (Aguilar & Plante, 2014; Von Koss Torkildsen et al., 2013) that linear order violations are easier to detect than co-occurrence violations. This can be explained by evidence that adjacent dependency learning is easier than nonadjacent dependency learning (Gómez, 2002; Gómez & Maye, 2005) and that adjacent dependency learning precedes nonadjacent dependency learning (Wilson et al., 2020). For example, in an aXd (NGC) violation, both the “aX” and “Xd” segments appear correct when analyzed through an adjacent-dependency lens. In contrast, in a bXa (NGO) violation, both the “bX” and “Xa” segments are incorrect according to adjacent rules, making this violation easier to identify. This explains why participants successfully detected linear violations but struggled with co-occurrence violations. It is also important to note that in this study, the three symbols of each string were presented all at once on the screen, which prevented an evaluation of sequential processing within and between strings. Future studies employing paradigms that emphasize sequential visual presentation are required to better understand participants' sensitivity to linear order violations observed in this study.

The integration of a follow-up test 1 week after the task as well as the ERP measurement provided valuable insights into the learning mechanisms of children with and without DLD. Observations from the follow-up test and ERP data suggest that children with TD engage in implicit rule extraction, leading to more robust learning that is supported by knowledge consolidation, while children with DLD may rely on rote-based associative learning, which leads to weaker learning and the gradual loss of learned information over time.

At the group level, the TD group maintained and consolidated their learning in the follow-up test, while the DLD group no longer exhibited a greater-than-chance performance at follow-up. This pattern aligns with findings from several consolidation studies in children with DLD (Desmottes et al., 2016, 2017; Hedenius et al., 2011), which reported that children with DLD did not exhibit the consolidation gains observed in TD groups during SL. Furthermore, by comparing two versions of a serial reaction time task with different practice session lengths, Desmottes et al. (2016, 2017) demonstrated that consolidation impairments in the DLD group persisted regardless of the training duration or the performance levels achieved during training. It is possible that children with DLD employed explicit learning strategies during the exposure phase and enjoyed an immediate positive effect on their test performance. The effect disappeared on the follow-up test a week later. This is consistent with the notion that knowledge acquired through implicit learning is more robust and stable than that rehearsed and memorized explicitly (Reber, 1989). This is in line with evidence that children with DLD tend to use rote memory and rely on explicit learning strategies for their learning detriment (Gopnik & Crago, 1991; Hsu & Bishop, 2010; Jones & Conti-Ramsden, 1997; Riches et al., 2006; Skipp et al., 2002).

At the individual level, 14 children with TD (50%) and only one participant with DLD (5%) exhibited clearer response consistency across all stimulus types in the follow-up test that was not observed in the immediate test, suggesting that they may have extracted a rule, whether partial or complete, to guide their judgments. This can be explained with the fuzzy trace theory (Brainerd & Reyna, 2002), which proposes that individuals encode the overall gist (a general event representation) as well as the verbatim details (the specific instance) for familiar events. Gist memory decays more slowly, and individuals may recall the gist without recalling verbatim information, making them less affected by the highly variable “X” elements in our task. Notably, all nine participants who successfully extracted both order and co-occurrence rules (i.e., complete rule extractors) belonged to the TD group, while only one participant with DLD demonstrated rule-guided judgment based on the partial rule extracted (i.e., detecting order violations but not co-occurrence violations). This suggests that children with TD were more likely to internalize and consolidate grammar-like rules over time, enabling them to apply learned rules systematically. In contrast, most children with DLD showed inconsistent patterns, suggesting that they may have relied on a trial-by-trial, probably familiarity-based judgment in the follow-up test. This highlights potential difficulties in encoding and consolidating abstract structures over time, which may contribute to persistent difficulties in grammar processing in this population. Interestingly, none of the complete rule extractors exhibited strong initial learning of co-occurrence rule during the immediate test, yet they successfully identified these violations in the follow-up test. This pattern implies that co-occurrence rule learning might require additional time for consolidation but does not necessarily require additional exposure.

Although not the primary focus of the study, exploratory analyses of individual differences revealed that language skills and age may play a role in predicting SL performance. Language skills measured by TILLS was a significant predictor of immediate test performance as well as long-term retention 1 week later. Age was a marginal predictor of immediate test performance but not for the follow-up test. These findings suggest that developmental and linguistic factors may differentially influence short- versus longer term learning outcomes. Short-term outcomes, measured by an immediate test performance, may involve more explicit strategies such as working memory, which continues to develop with age (Gathercole et al., 2004; Luna et al., 2004). Long-term outcomes, on the other side, may depend more heavily on automatic consolidation processes associated with implicit SL, which was found uniformly effective until the age of 12 years (Janacsek et al., 2012). However, given the modest sample size and exploratory nature of these analyses, these interpretations should be viewed as preliminary and warrant further investigation in future research.

Our ERP analysis focused on components commonly associated with language processing, the P600 and N400. Although the P600 is not uniquely a “syntactic” component, in the context of grammaticality judgments, it provides insight into the extent to which grammatical rules were learned: Grammatical violations are treated as deviant events that prompt updating of the internal representation of the environment. Our findings revealed a P600 response elicited by linear order violations only in the TD group, suggesting that TD children engaged in syntactic integration and reanalysis (P600) that may be critical for learning and generalization. The absence of this response in children with DLD may reflect challenges in higher order syntactic processing and could help explain their decline in accuracy at follow-up.

We did not observe a classic centro-parietal N400 component in response to grammatical violations in either group. Instead, both linear order violation and co-occurrence violations elicited a frontal negativity peaking around 400 ms at Fz in children with DLD, resembling the FN400 component commonly reported in recognition memory literature. Neither violation type elicited such effect in TD children. The FN400 component has traditionally been associated with familiarity-based recognition in episodic memory tasks, where it manifests as a frontal negativity to unfamiliar or less fluent stimuli (Curran, 2000; Rugg & Curran, 2007; Voss & Federmeier, 2011). In this context, the FN400 effect observed for both types of grammatical violations in children with DLD may reflect a fluency disruption response to less familiar stimuli. It is possible that children with DLD processed these strings as a single unit that was either familiar or unfamiliar to them based on the initial exposure. This finding suggests that children with DLD may have engaged rote memory during the initial exposure. The presence of the FN400 in response to ungrammatical strings in children with DLD may suggest that they may have detected some surface-level patterns, not necessarily underlying rules in the immediate test. Such effect was not fully reflected in their behavioral performance, implying that although these strings were unfamiliar, they were not necessarily judged as “incorrect” explicitly.

The observation of the FN400 only in the DLD group suggests a potentially compensatory or alternative processing mechanism in this group. Whereas TD children may have formed more abstract grammatical representations, children with DLD may have relied more heavily on item-level familiarity or pattern-based fluency. This interpretation aligns with theories proposing that children with DLD may rely on declarative memory to support the detection of irregularities or disruptions in fluency, but may struggle with rule abstraction and consolidation (Courteau et al., 2023; Gul et al., 2023). The frontal topography and timing of the observed effect support the idea that children with DLD may have processed grammatical violations through mechanisms associated with familiarity-based detection, rather than syntactic integration. Our findings also contribute to ongoing debates about the relationship between the FN400 and the classic N400. While some researchers have proposed that the FN400 is a frontal variant of the N400 (Paller et al., 2007; Voss & Federmeier, 2011), the current data indicated that FN400 can be elicited within a semantic-free artificial grammar learning task, supporting its sensitivity to unfamiliarity or disfluency detection.

Taken together, these findings suggest that children with DLD relied on item-based familiarity detection (FN400) rather than structured syntactic processing and integration (P600). While they may have picked up on statistical regularities at a surface level and experience disruptions when those regularities are violated, their learning did not translate into rule abstraction and long-term consolidation, possibly due to a greater reliance on verbatim memory rather than structural generalization. This reliance on familiarity-based processing instead of syntactic integration may contribute to persistent difficulties in acquiring and consolidating grammatical structures in children with DLD.

Conclusions

In summary, the P600 sensitivity to syntactic violations found only in the TD group may represent a rule-governed, top-down process, while the frontal negativity (FN400) found only in children with DLD may suggest a process of rote-based associative learning. These findings are supported by the group differences that emerged in the follow-up test 1 week after learning, and the findings that rule extractors were primarily from the TD group. The follow-up test results revealed that the TD group consolidated their learning, whereas the DLD group did not. While 14 out of 28 children with TD demonstrated rule-based discrimination, only one out of 20 children with DLD exhibited this pattern. Our findings also provide support to the feature deficit hypothesis by Gopnik and Crago (1991), which suggested that children with DLD might treat morphological markers as random phonological variants and learn by rote. These findings suggest that even when children with DLD achieve equivalent immediate learning outcomes as their TD peers do, they might still need additional support for long-term knowledge consolidation. Future research should explore and evaluate different educational strategies to address this need.

Limitations and Future Directions

While this study provides valuable insights into SL in children with and without DLD, several limitations should be acknowledged. First, the sample size was relatively small, particularly for the DLD group (n = 22), which may have limited statistical power to detect more subtle effects. Future studies should aim to replicate these findings with larger samples to enhance generalizability. Second, our task design relied solely on visual stimuli, which, while effective in isolating SL mechanisms, leads to limited ecological validity and may not fully capture the learning difficulties faced by children with DLD in real-world contexts. Incorporating multimodal learning conditions with auditory or other paradigms could provide a more comprehensive understanding of how these children process linguistic regularities.

Third, the design of the follow-up test was imperfect. EEG data were not collected during the remote follow-up test, preventing direct comparisons of neural processing between immediate and delayed learning outcomes. Additionally, the test environment and the number of trials in the immediate test were different from the follow-up test (remote with fewer trials). Future studies should consider an in-person follow-up test using identical procedures to the immediate test.

Finally, the fixed 1-week interval used in this study precludes an examination of how SL and knowledge consolidation unfold over different timescales. Future research should manipulate test intervals ranging from days to months to better understand the trajectory of learning consolidation in children with and without DLD. By addressing these limitations, future studies can further elucidate the mechanisms underlying SL in DLD and inform the development of more effective language intervention strategies tailored to this population.

Data Availability Statement

The data sets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.

Appendix

Figure A1.

Topographic maps of difference (NGC − G, NGO − G) in each group (typical development [TD], developmental language disorder [DLD]) in the P600 time window (left) and FN400 time window (right). G = grammatical; NGC = co-occurrence violation; NGO = linear order violation.

Funding Statement

This material is based on work supported by the National Institute on Deafness and Other Communication Disorders under Grant R01DC018295 awarded to Yael Arbel.