Seven-year-olds recall non-adjacent dependencies after overnight retention
Introduction
Language is made up of different building blocks, combined together to form sentences. The grammar of a given language defines the rules for these combinations. For example, grammatical rules define that determiners can be combined with nouns (The girl), but not with verbs (*The give). Grammatical dependencies can be formed not only between neighboring elements, but also between non-neighboring elements of a sentence. For example, in the sentence The girlSg smilesSg, girlSg and -sSg form a grammatical dependency (i.e., number agreement) that spans one element (smile). In theory, these dependencies can span an arbitrary number of elements, as demonstrated in the following example: The girlSg who visited us yesterday smilesSg. These types of dependencies, called non-adjacent dependencies (NADs), are important grammatical rules of a language, such that becoming a proficient speaker and listener of languages highly depends on acquiring these rules (see Wilson et al., 2018).
Adults have been shown to be able to process and learn NADs in a number of behavioral studies (e.g. Frost and Monaghan, 2016, Gómez, 2002, Newport and Aslin, 2004, Peña et al., 2002). For example, Gómez (2002) exposed adults to an artificial language containing NADs in the form of three-syllable strings. In this study, the artificial language learning task consisted of NADs that were realized as AXC structures, with A and C being the dependent elements and X being variable elements. After familiarization to these strings, participants were shown a mixture of strings, either containing familiarized NADs or NAD violations; they were asked to indicate whether a given string followed the rules of the familiarized artificial language. The results showed that adults are in principle able to learn NADs (Gómez, 2002). However, adults’ NAD learning has been shown to be somewhat restricted, as several studies demonstrated that adults only successfully learned NADs when phonological cues between dependent elements were provided (Mueller et al., 2008, Newport and Aslin, 2004, Peña et al., 2002). Taken together, behavioral studies demonstrated that adults are able to learn NADs in an artificial language. Becoming a successful speaker, however, already starts in early infancy; and even infants have been shown to be able to learn NADs (Gómez and Maye, 2005, Lany & Gómez, 2008). For example, Gómez and Maye (2005) exposed infants to AXC grammars using the Head Turn Preference Procedure (Nelson et al., 1995), with which they measured infants’ looking time towards an auditory stream played on either side of the infant. Specifically, they first familiarized infants with the AXC grammar (i.e., NADs), which was then followed by the presentation of correct or incorrect (i.e., containing a violation) NADs. Fifteen-month-old infants oriented more towards the familiarized stimuli (i.e., correct NADs) than to violations (i.e., incorrect NADs), indicating that 15-month-olds learned the AXC grammar (Gómez & Maye, 2005). However, infants’ NAD learning underlies some restrictions depending on how the NADs are presented (e.g. Höhle, Schmitz, Santelmann, & Weissenborn, 2006, Santelmann & Jusczyk, 1998). For example, 18- to 19-month-old infants can only learn NADs when the intervening elements consist of three syllables or less. If there are more intervening elements, NAD learning breaks down (Höhle, Schmitz, Santelmann, & Weissenborn, 2006, Santelmann & Jusczyk, 1998). Taken together, both adults and infants are able to learn NADs in principle. However, the processes underlying NAD learning cannot be fully understood by using offline behavioral methods alone, but should be supplemented by online methods, such as the serial reaction time task or the click detection task (Gómez, Bion, & Mehler, 2011, Misyak, Christiansen, & Tomblin, 2010). In addition, event-related potentials (ERPs) have been used as an online method to investigate NAD learning. Such online methods allow the more direct examination of the time course of learning and the possible change of underlying learning mechanisms.
Mueller, Oberecker, and Friederici (2009) used ERPs to investigate the learning of NADs that were embedded in natural speech in a foreign language (Italian). During familiarization, they exposed German native speakers, without prior knowledge of Italian, to Italian sentences containing NADs (e.g. “La sorella sta cantando”; the sister is singing). In testing phases, participants heard a mixture of correct sentences and incorrect sentences containing NAD violations (e.g. “La sorella sta cantare”; the sister is sing∅). By comparing ERPs to incorrect sentences with ERPs to correct sentences during testing phases, a series of studies could show that both infants (under passive listening conditions, i.e. without a task; Friederici, Mueller, & Oberecker, 2011) and adults (under active conditions, i.e. with a task; Mueller et al., 2009) are able to learn these NADs embedded in a miniature version of Italian. Infants showed a more positive ERP response to incorrect compared to correct NADs, while adults showed a more negative ERP response. Interestingly, when adults’ prefrontal cortex (PFC) was inhibited using transcranial direct current stimulation (tDCS), adults’ ERP response to incorrect compared to correct NADs changed from a negative ERP to a late positive ERP, which was interpreted to indicate different underlying processes (Friederici, Mueller, Sehm, & Ragert, 2013). The late positive ERP found in adults whose PFC was inhibited was similar to infants’ positive ERP responses to NAD violations (Friederici et al., 2011), whose PFC is not yet fully developed (Huttenlocher, 1990). Thus, the polarity difference of the ERP responses to NAD violations seems to be not only due maturational changes between infancy and adulthood, but moreover due to an underlying difference in learning mechanisms. Similarly, studies of language development in early childhood have demonstrated that the polarity of an ERP effect and a developmental change of the ERP effect polarity can be meaningful in terms of later behavior and indicative of different underlying processes (Kooijman et al., 2013, Schaadt et al., 2015; see also Eimer, Forster, & Van Velzen, 2003, and Penney, Mecklinger, & Nessler, 2001, for evidence of a reversal of polarity that is indicative of behavior in adults).
As indicated by a difference in ERP polarity of components elicited by NAD violations, infants and adults might use different learning mechanisms and develop different representations of the NADs. Specifically, it has been suggested that infants learn NADs more automatically than adults do, also reflected in infants’ ability to learn under passive listening, which adults struggle to do (Mueller, Friederici, & Männel, 2012). Interestingly, Mueller, Friederici, and Männel (2018) showed that children up to the age of 2 years are able to learn NADs under passive listening conditions, while 4-year-olds, similar to adults (Mueller et al., 2012) struggle to do so and may need active task conditions. It has been suggested that the specific need for an active task is associated with a switch in learning mechanisms from associative, bottom-up learning (allowing learning under passive listening conditions) to controlled, top-down learning (hindering learning under passive listening conditions, but facilitating learning under active task conditions; see Skeide & Friederici, 2016). This switch may be associated with PFC maturation (Skeide & Friederici, 2016), which reaches near adult-like maturity around the age of 7 years (Huttenlocher, 1990). While this claim has not been tested longitudinally, there is some evidence for this from the tDCS (Friederici et al., 2013) and the cross-sectional study (Mueller, Friederici, & Männel, 2018) described above. Taken together, NAD learning mechanisms change during development, which may possibly be linked to PFC development.
Although previous studies (Friederici et al., 2011, Mueller et al., 2009, Mueller et al., 2012) convincingly demonstrated that individuals can differentiate familiarized NADs from NAD violations, NAD learning was always tested on the day of the familiarization itself, either on the same items as during familiarization (e.g., Friederici et al., 2011, Mueller et al., 2009) or on novel items (i.e. items that share the same structure as familiarized items, but use different tokens; e.g., Gómez, Bootzin, & Nadel, 2006). This testing procedure provides a measure of whether participants have formed a representation of the familiarized items, which can then be compared to test items. Test items perceived as similar to familiarization items would then be interpreted as adhering to the (possibly unknown) rule. On the other hand, test items judged as dissimilar would be interpreted as not adhering to the underlying rule (similarity-based learning; see Opitz & Hofmann, 2015). While testing NAD learning on the same day of familiarization is certainly informative, it is a matter of discussion whether this should be interpreted as evidence that the underlying rules have been learned, rather than some surface-based features of the NADs. This is because the knowledge of the underlying rules that characterize the (artificial) grammar only builds up over time (Opitz & Hofmann, 2015) and might not be fully present immediately after a relatively brief familiarization. Thus, it is important to retest NAD learning after a period of time in order to investigate whether NAD learning had a lasting effect and to show that learned NADs are not simply forgotten again shortly after familiarization. In order to investigate whether participants have really learned the underlying rules and could recall them after a period of time, several studies have investigated recall of grammatical rules after a retention period. For example, Fischer, Drosopoulos, Tsen, and Born (2006) investigated the effect of a retention period on artificial grammar learning in adults. The authors showed that before sleep there was no evidence for above-chance level performance in adults in a generation task, during which participants had to predict the next letter in a string based on the artificial grammar. However, after a retention period involving sleep participants could solve the task successfully, which was not the case after a retention period without sleep. A number of studies has demonstrated that this benefit of a retention period involving sleep is linked to a change in representations (see Diekelmann and Born, 2010, Ellenbogen et al., 2007, Fischer, Drosopoulos, Tsen, & Born, 2006, Wagner et al., 2004). Davis and Gaskell (2009) suggested that a model of memory consolidation, the complementary learning systems model, could also apply to the linguistic domain, specifically word learning. Under this model, new knowledge, such as a newly encountered word, is initially stored in episodic memory, where it is not yet integrated into the lexicon. New words are then consolidated into lexical memory over time, facilitated by sleep (Henderson, Weighall, Brown, & Gaskell, 2012, Smith et al., 2018, Tamminen et al., 2010). Especially infants and children were shown to benefit from a retention period (particularly when retention involved sleep; Backhaus et al., 2008, Henderson, Weighall, Brown, & Gaskell, 2012, Hupbach et al., 2009, Smith et al., 2018) and for generalizing learned information to new input (Gómez et al., 2006). A study by Friedrich, Wilhelm, Mölle, Born, and Friederici (2017) linked a change in representations of learned associations during the course of a retention period to particular ERPs. In this study, infants were exposed to object-word pairs followed by a retention period that either involved a long nap, a short nap, or no sleep. Before retention, there was no evidence for learning of the object-word pairs and neither did the group without sleep show any sign of learning after retention. In contrast, infants who had a short retention period (30 min on average) involving sleep showed consolidation of the object-word pairs. However, the ERPs only revealed a late negativity, which was interpreted to be indicative for a phonological association between the word and object, but not for a lexical-semantic representation of the object-word pairs in long-term memory. Only those children who had a longer consolidation period (50 min on average) involving sleep also showed ERP evidence of lexical-semantic representations of word meaning in long-term memory in form of an N400 (i.e., earlier negativity; Friedrich et al., 2017). Thus, this study demonstrates that children benefit from a retention period involving sleep, which most likely leads to the ERP effects of successful recall of learned associations after the retention period. Given these promising findings showing a beneficial effect of a retention period involving sleep on long-term memory consolidation, we aimed at investigating the effect of retention involving sleep on the recall of NADs as important grammatical rules of language.
Thus, in the present ERP study, we investigated 7-year-old children’s recall of NADs embedded in a miniature version of a foreign language (i.e., Italian), using the same paradigm as Mueller et al. (2009), including a grammaticality judgment task. We invited our participants on two consecutive days, ensuring a retention period involving sleep to test recall of NADs. If we can show recall of NADs on day two, we provide evidence that children learned the NADs and that this learning had lasting effects beyond the familiarization day, which goes over and above showing processing differences between correct and incorrect NADs on the same day when familiarization took place. We tested 7-year-olds because they have been shown to be able to successfully perform offline behavioral tasks assessing statistical learning (Raviv and Arnon, 2017, Shufaniya and Arnon, 2018), most likely associated with 7-year-olds’ advanced PFC maturation (Huttenlocher, 1990), playing a crucial role in NAD learning (Friederici et al., 2013).
A number of recent studies have raised concerns that group-level offline tasks, which assess statistical learning, may not provide reliable measures of individual differences (Siegelman et al., 2017, West et al., 2018), particularly in children (Arnon, 2019). Siegelman, Bogaerts, Christiansen, and Frost (2017) suggest that online measures may circumvent some of the problems seen in the reliability of offline tasks. Here, we use ERPs as an online test of NAD learning both at the group level and the individual level, as ERPs have been shown to be a reliable measure of interindividual differences in a variety of paradigms (Cassidy, Robertson, & O’Connell, 2012).
According to the procedure of Mueller et al. (2009) in adults, children listened to only correct stimuli (i.e., Italian sentences) during the four learning phases on the first testing day. Each learning phase was followed by a testing phase, during which children listened to incorrect stimuli containing NAD violations intermixed with correct stimuli following the familiarized NAD rule. During the testing phases, children were required to behaviorally indicate whether or not a given stimulus belonged to the language they were familiarized with in the learning phases (i.e., grammaticality judgment task). On the following day, we tested recall of NADs by asking children to perform only the four testing phases, again including the grammaticality judgment task. To capture consolidation and recall of NADs on the next day, we specifically focused on the change in behavior from day one to day two. Successful recall of NADs will be reflected in behavioral improvement from day one to day two (i.e., more correct grammaticality judgments on day two compared to day one). If children learn the NADs on day one and recall them on day two, we expect that children’s ERP responses on both days are associated with their improvement in the number of correct grammaticality judgments from day one to day two. While we will treat this correlational analysis as an exploratory analysis due to reliability concerns (see Siegelman, Bogaerts, Christiansen, et al., 2017), linking ERPs to the behavioral outcome may strengthen the interpretability of our results.
Section snippets
Participants
For the present experiment, 49 children were invited. The datasets of 36 children (20 boys) with a mean age of 7.22 years [Standard Deviation (SD) = 0.36] entered the final analyses (i.e., the datasets of 13 children were excluded due to movement and perspiration artifacts in the EEG). Children visited the first and second school grade. All participants were German monolinguals and none of the children had any known hearing deficits or neurological problems. In order to ensure that the Italian
Behavioral results
Correct answers in percent on day one (mean = 48.65%; SD = 5.90) did not differ significantly from correct answers on day two (mean = 50.73%; SD = 6.18; t (35) = −1.49; p = .15). RTs were significantly shorter on day two (mean = 7489.13 ms; SD = 3843.89) compared to day one (mean = 11552.69 ms; SD = 4847.99; t (35) = 5.77; p < .001). When using the criterion of the binomial test (i.e., when z was set to 1.64, resulting in a above-chance level threshold of 58.2% correct answers in percent), we
Discussion
The aim of the present ERP study was to investigate NAD learning by means of NAD violation recall using a miniature version of a natural language in 7-year-olds. Specifically, we not only tested NAD processing directly after learning, but also after a retention period involving sleep (i.e., at the next day). On the first day, German-speaking children were exposed to Italian sentences containing NADs (e.g. La sorella sta cantando; the sister is singing). Learning phases were followed by testing
CRediT authorship contribution statement
Gesa Schaadt: Formal analysis, Writing - original draft, Visualization. Mariella Paul: Writing - original draft. R. Muralikrishnan: Conceptualization. Claudia Männel: Writing - review & editing, Supervision. Angela D. Friederici: Conceptualization, Writing - review & editing, Supervision.
Acknowledgements
The Max Planck Society (GS, MP, RM, CM, ADF), the German Research Foundation, project FR 519/20-1 (FOR 2253; GS, MP, CM, ADF), and the Berlin School of Mind and Brain (MP) funded this project. The authors declare to have no conflict of interest. Further, we would like to thank all the participating families for their commitment, as well as Sophia Röthing for her help with setting-up the experiment and Ulrike Barth for her help with data acquisition and her dedicated work with the participating
References (60)
- et al.
Mass counts: ERP correlates of non-adjacent dependency learning under different exposure conditions
Neuroscience Letters
(2011) - et al.
Syntactic comprehension in Parkinson's disease: Investigating early automatic and late integrational processes using event-related brain potentials
Neuropsychology
(2003) - et al.
The sleeping infant brain anticipates development
Current Biology
(2017) - et al.
Simultaneous segmentation and generalisation of non-adjacent dependencies from continuous speech
Cognition
(2016) - et al.
Similarity and rules: Distinct? exhaustive? empirically distinguishable?
Cognition
(1998) - et al.
The N400 effect in children: Relationships with comprehension, vocabulary and decoding
Brain and Language
(2011) Morphometric study of human cerebral cortex development
Neuropsychologia
(1990)- et al.
Electrophysiology reveals semantic memory use in language comprehension
Trends in Cognitive Sciences
(2000) - et al.
Learning at a distance I. Statistical learning of non-adjacent dependencies
Cognitive Psychology
(2004) - et al.
Interactions of the hippocampal system and the prefrontal cortex in learning language-like rules
NeuroImage
(2003)
Concurrence of rule-and similarity-based mechanisms in artificial grammar learning
Cognitive Psychology
Repetition related ERP effects in a visual object target detection task
Cognitive Brain Research
Spherical splines for scalp potential and current density mapping
Clinical Neurophysiology
Sensitivity to discontinuous dependencies in language learners: Evidence for limitations in processing space
Cognition
Present and past: Can writing abilities in school children be associated with their auditory discrimination capacities in infancy?
Research in Developmental Disabilities
Do current statistical learning tasks capture stable individual differences in children? An investigation of task reliability across modality
Behavior Research Methods
Immediate as well as delayed post learning sleep but not wakefulness enhances declarative memory consolidation in children
Neurobiology of Learning and Memory
Retest reliability of event-related potentials: Evidence from a variety of paradigms
Psychophysiology
A complementary systems account of word learning: Neural and behavioural evidence
Philosophical Transactions of the Royal Society B: Biological Sciences
The memory function of sleep
Nature Reviews Neuroscience
Anterior and posterior attentional control systems use different spatial reference frames: ERP evidence from covert tactile-spatial orienting
Psychophysiology
Human relational memory requires time and sleep
Proceedings of the National Academy of Sciences
Implicit learning - explicit knowing: A role for sleep in memory system interaction
Journal of Cognitive Neuroscience
Precursors to natural grammar learning: Preliminary evidence from 4-month-old infants
PLoS One
Language learning without control: The role of the PFC
Journal of Cognitive Neuroscience
Variability and detection of invariant structure
Psychological Science
Naps promote abstraction in language-learning infants
Psychological Science
The word segmentation process as revealed by click detection
Language and Cognitive Processes
The developmental trajectory of nonadjacent dependency learning
Infancy
Brain signatures of syntactic and semantic processes during children’s language development
Journal of Cognitive Neuroscience
Cited by (0)
- 1
These authors contributed equally.