Do children outperform adults in implicit language acquisition? Evidence from a statistical learning paradigm

(1)

DO CHILDREN OUTPERFORM ADULTS

IN IMPLICIT LANGUAGE ACQUISITION?

EVIDENCE

FROM

A

STATISTICAL

LEARNING PARADIGM

.

Word count: 21,666

Maya Braun

Student number: 01805814

Supervisor(s): Dr. Eleonore Smalle

A dissertation submitted to Ghent University in partial fulfilment of the requirements for the degree of Master of Clinical Psychology

(2)

Preamble concerning COVID-19

Because this thesis was written in the midst of the global corona pandemic in 2020, the author had to adhere to the corona restrictions installed by the Belgian government. Since data collection was already completed before the start of the restrictions, the present thesis did not suffer from the measures taken. This preamble was drafted by the student in cooperation with the promotor and confirmed by both.

Corona preambule

Omdat deze masterproef geschreven werd tijdens de globale coronacrisis in 2020 moest de auteur zich tijdens het schrijven houden aan de corona-maatregelen van de Belgische regering. Omdat de dataverzameling al voor het begin van de maatregelen afgerond was, heeft het uitwerken van deze masterproef niet geleden onder de maatregelen. Deze preambule werd in overleg tussen de student en de promotor opgesteld en door beiden goedgekeurd

(3)

Abstract

While children are long assumed to be better language learners than adults, theories on maturational effects for language acquisition are heavily debated. Today, the reasons why children outperform adults in language learning remain unclear. Speech segmentation is one of the earliest steps in early language acquisition on which children are thought to outperform adults. It refers to the process of identifying novel words in continuous speech via statistical cues, also defined as statistical learning. The present paper unites theories about memory with theories about language learning in order to gain a better understanding of language acquisition mechanisms throughout the lifespan. To do so, a total of 50 children aged 7 to 11 years and 50 young adults completed measures for learning on a speech segmentation task. The aim was to separate implicit and explicit memory components of language acquisition for different age groups. In the first task, children showed significant better implicit knowledge of the words than adults. In the second task, children acquired the same amount of statistical knowledge faster than adults. Eventually, both groups achieved the same level of learning. The different measures of learning correlated significantly with each other, indicating that they measure the same learning construct. Overall, the results of the current study indicate that children outperform adults in implicit measures of artificial language acquisition. These findings suggest that the language learning advantage of children is due to a competitive interaction between implicit and explicit memory.

(4)

Samenvatting

Ook al wordt aangenomen dat kinderen beter zijn dan volwassenen in het leren van taal, blijven theorieën rond het effect van cognitieve maturiteit op taalverwering sterk omstreden. Tot op vandaag zijn de redenen waarom kinderen volwassenen zouden overtreffen in taalverwerving nog steeds onduidelijk. Spraaksegmentatie is één van de eerste stappen in de vroege taalverwerving waarvan wordt vermoed dat kinderen het beter doen dan volwassenen. Dit betreft het proces van het identificeren van nieuwe woorden binnen een continue stroom van statistische cues, het zgn. statistisch leren. Deze masterproef verenigt theorieën uit geheugenonderzoek met kennis over taalverwerving om een beter inzicht te krijgen in taalverwervingsmechanismen doorheen het leven. Hiervoor legden 50 jonge volwassenen en 50 kinderen tussen 7 en 11 jaar oud testen van spraaksegmentatie af. Het doel was om impliciete en expliciete geheugencomponenten van taalverwerving voor de twee leeftijdsgroepen te onderscheiden. In de eerste taak toonden kinderen dat ze een significante hoeveelheid impliciete kennis konden bijhouden, terwijl volwassenen dit niet konden. In de tweede taak leerden kinderen dezelfde hoeveelheid informatie sneller dan volwassenen. Uiteindelijk bereikten beide groepen hetzelfde niveau van leren. De resultaten van de twee taken correleerden significant met elkaar en lijken dus hetzelfde construct te meten. Dit onderzoek toont aan dat kinderen volwassenen overtreffen m.b.t. impliciete maten van taalverwerving. Deze bevindingen suggereren dat het taalleervoordeel van kinderen komt door een competitieve interactie tussen impliciet en expliciet geheugen.

(5)

Language is considered one of the most defining traits of humanity in its evolutionary process (Burnet, 1774; Friederici, 2017). Without it, our complex social systems would not be able to function the way that they do now. In current day society, globalisation brings people of different nations and mother tongues closer together. To work together despite such a language barrier, the acquisition of language becomes increasingly important. It has become a dominant topic in education all over the globe. Speaking multiple languages increases professional opportunities, but also enables access to a greater variety of cultural resources. It allows having social contacts across different countries, therefore broadening our social networks. Thus, language acquisition has implications for the professional, the cultural, and the personal lives of language learners (Cenoz et al., 2006).

Many claim that languages are best learned early in life, with some well-established researchers claiming that they cannot be learned beyond a specific age (Lenneberg, 1967). Therefore, both parents and schools put an increasing amount of effort in teaching languages at the youngest age possible. On the other hand, older learners often feel discouraged at learning a new language. But does this common understanding of language learning throughout the lifespan reflect reality?

In the introduction of this thesis, the critical period for language learning will be discussed along with other prominent theories for learning and its development across the lifespan. More particularly, existing theories and paradigms from memory and language acquisition research, and also present promising evidence on human motor skill learning, will be presented. The aim of this paper is to answer questions that neither of those disciplines could answer by themselves: Do specific age groups have an advantage on language learning, and if so, why?

Sensitive Age Hypothesis in Language Acquisition

Despite the high level of interest in the study of language, many questions remain unanswered. One of them concerns the question of children having an advantage over adults when it comes to learning a language. There seems to be a strong consensus between researchers that children up to a certain age outperform adults in language learning, whether it concerns a first or second language (Kennedy & Norman, 2005; Lenneberg, 1967; Newport, 1994; Smalle et al., 2016). However, there is no agreed upon explanation why children should

(7)

have this advantage. One of the first theories that proposed such early advantage in language learning was Lenneberg’s critical period hypothesis (1967). According to the critical period hypothesis, one’s native language has to be learned before the critical age of twelve, i.e. before the start of puberty, or otherwise full mastery of the language will never be obtained. Lenneberg (1967) argued that this was because the language lateralisation process of the brain ends before puberty. However, this hypothesis received some criticism. Firstly, empirical research found that the so-called age-dependent decrease in language learning performance (i.e., after puberty) is gradual rather than abrupt, making the existence of a critical age too conservative (Moskovsky, 2001). Secondly, Lenneberg based his hypothesis mostly on children who were deaf, feral or had major cognitive impairments (Vanhove, 2013), and were thus not representative for the general public.

Lenneberg’s hypothesis was later applied to second-language acquisition in empirical research (Johnson & Newport, 1989). For instance, Johnson and Newport (1989) studied language learning in native Korean and Chinese speakers who immigrated to the United States at different ages. The authors compared their age upon arrival with English language learning ability. To test the age disparity, participants were asked to take an English grammar judgement task. The authors found a linear relationship between English language skills and age-upon-arrival, with a younger age-upon-arrival being associated with higher performance on the grammar judgement test. This correlation was present for ages before puberty, but not after, even after controlling for prior experience with English, motivation, self-consciousness and how strongly they identified with American culture. These findings indicate that age plays an important role in second language acquisition, but the decrease around puberty seems to be less abrupt than was initially suggested by Lenneberg (1967). Adult immigrants were still able to acquire the language, although not to the same proficiency level as children.

The differentiation between first and second language acquisition is essential. Some authors argue that there is a critical period for the acquisition of the first language after which the ability to acquire language, more specifically its grammatical system, is lost (Moskovsky, 2001). For second language learning, it seems likely that there is a sensitive period for language learning, rather than a critical one (Oyama, 1976; Vanhove, 2013). Authors argue that adults can still acquire a second language, though not to the same proficiency as young learners (Moskovsky, 2001). The gradual nature of the decrease of second language learning as a function

(8)

of age, as observed in Johnson and Newport (1989), contradicts the existence of a critical period and corresponds more to a sensitive period hypothesis (Moskovsky, 2001).

Another important observation in the study of language learning is that the effect of age seems to vary as a function of the linguistic aspect one is studying. While new vocabulary and semantic processing can be mastered just as well by older learners as by younger ones (Johnson & Newport, 1989), complex grammar and phonetic learning seem to be more influenced by age (Sanders, Yamada, & Neville, 1999).

In sum, the findings on a critical age effect in language learning are inconsistent. While some researchers find it only for first language acquisition (Vanhove, 2013), others also report comparable age effects for the acquisition of a second language (Johnson & Newport, 1989). Furthermore, some researchers could only find an age effect for specific aspects of language learning, such as phonological and grammatical structures (Newport, Bavelier, & Neville, 2001). As a result, the main questions that remain mostly unanswered today are the following: Do children outperform adults on second language learning, and if so, why? The question of the “why” was even pronounced as one of the most important research questions in the 125th

-anniversary issue of the magazine Science (Kennedy & Norman, 2005). This phenomenon is especially surprising when considering that adults outperform children in many other measures of cognitive functioning, such as factual learning, attentional control and working memory, due to a late maturation of the prefrontal cortex (PFC, Craik & Bialystok, 2006). If children’s cognitive system has not fully matured yet, it would be expected that they are outperformed by adults in every domain of learning. Then why do they learn languages seemingly effortlessly?

The Role of Memory in Language Learning

Language learning has been a distinct field of interest for a long time. While this allowed for focused research to happen on this specific skill, it did not encourage researchers to look at approaches in different fields of studies that could be applied to language learning. As a result, language learning literature has co-existed with other learning literature without many attempts to apply these models to language.

The present paper will explore theories from memory research. Since memory is a crucial aspect of any kind of learning, applying it to language acquisition can provide further insight. By combining implications of two models of memory, a first possible explanation for the

(9)

language learning advantage can be concluded. In short, the Declarative/Procedural Model divides memory into two systems and categorizes different aspects of language acquisition accordingly. The Competing Memory Systems approach then offers a framework that can potentially explain the differences between adults and children when it comes to language acquisition.

The Declarative/ Procedural Model.

A well-researched model of the role of memory in language is the Declarative/ Procedural model from Ullman (2001). This model assumes that there are two memory systems involved in learning languages: the procedural memory system and the declarative memory system. Both of those systems are assumed to store information on the long term.

The procedural system is responsible for the automatisation of actions and receives information without conscious reflection. The result of these processes is implicit learning. Implicit learning is defined as learning without awareness or intention, which results in knowledge that is difficult to express verbally (Cleeremans, Destrebecqz, & Boyer, 1998). Common examples of implicit or procedural learning stem from motor learning, such as learning how to ride a bicycle. When we ride a bicycle, we do not consciously think about what we do to move forward. Instead, we know the movements and do them automatically. Learning in the procedural system is reflected by changes in the neural circuits that were activated during the learning process. This includes processing areas such as the basal ganglia, the cerebellum and parts of the neocortex (Reber & Squire, 1994; Squire, 2004). These areas of the cortex reach full maturity early in life (i.e., before puberty) (Ullman, 2001). Learning in this system happens best via repeated exposure (Ullman, 2001; Ullman & Lovelett, 2018). In our example of riding a bicycle, this means that repeated practice is needed before full mastery of the process is obtained.

The declarative system, on the other hand, encodes and retrieves information via explicit association (Lum & Kidd, 2012). Semantic information and episodic events are stored in this system (Lum & Kidd, 2012). An example of a declarative learning process is studying for theoretical exams. It requires active, explicit revision of the study material, and results in conscious, retrievable knowledge. The declarative system is rooted in medial temporal areas, including the hippocampus, and frontal areas which reach maturity in young adulthood (Reber

(10)

& Squire, 1994; Squire, 2004). Learning in this system happens quickly, with only comparatively little exposure being needed (Ullman, 2001; Ullman & Lovelett, 2018).

The Two Memory Systems and Language Learning.

Researchers found that the declarative and procedural memory systems contribute distinctively to language learning (Ferman, Olshtain, Schechtman, & Karni, 2009; Ullman, 2001). This theory applies to both first and second language acquisition. The vocabulary seems to mostly depend upon declarative memory (Ullman, 2001). Words and their meanings are memorised explicitly, and their acquisition results in conscious knowledge. The same goes for irregular grammar (Ullman, 2001). However, regular grammar, phonology and syntax seem to depend mostly on procedural memory (Saffran, 2002; Ullman, 2001). Ullman (2001) argues that grammar learning and phonology is computational, meaning that its acquisition is based on the statistical likelihood of the input. For example, when exposed to the English grammar, we unconsciously learn that verbs that follow on he, she or it usually end in an “s”. This knowledge is formed because it is a rule that regularly applies. A verb ending in an s after he, she or it is therefore much more likely to occur than a verb that is not ending in an s. Our brain starts to predict the s subconsciously, and every confirmation of this rule strengthens this knowledge.

This is learned without being able to express those rules explicitly, as observed within native speakers who are confronted with grammar questions about their language (Ullman, 2001). More often than not, they find themselves just knowing how to say things, or having a

feeling for it. This intuition is a procedural aspect of language learning. Importantly, this is only

true for regular grammar. Irregular grammar, such as irregular verb forms in English, is not learned this way. There are different kinds of evidence for this claim. Firstly, the acquisition of these different language structures matches the corresponding patterns of learning of procedural or declarative memory. Practically, this means that language aspects covered by procedural memory, such as regular grammar, gradually improve with repeated exposure (Ullman, 2001; Ullman & Lovelett, 2018). For phonology, a well-studied learning process concerns speech segmentation (Saffran, Newport, & Aslin, 1996). This refers to the ability to detect borders between words and therefore perceiving a sentence not just as one big stream of sounds, but as a combination of distinct units (words). Speech segmentation will be described more elaborately in the statistical learning section of this paper. Since this process concerns

(11)

procedural learning, speech segmentation improves gradually with repeated exposure (Ullman, 2001; Ullman & Lovelett, 2018).

Language aspects that rely on the declarative system, such as irregular grammar and vocabulary, do not seem to improve significantly with repeated exposure. Evidence for this can be found for irregular past tenses in English (Ullman, 1999, 2001) and irregular plural in German (Clahsen, Eisenbeiss, & Sonnenstuhl-Henning, 1997; Penke & Krause, 2002).

Further evidence for the role of the different memory systems in language learning can be found when looking at clinical populations. Patients with neural damage on the basal ganglia, associated with the procedural system, showed significantly impaired performances in corresponding language tasks, such as the use of regular past tense (Ullman et al., 1997; Ullman et al., 2005). Interestingly, those patients also tended to show deficits in motor sequences (Vargha-Khadem, Watkins, Alcock, Fletcher, & Passingham, 1995), which are assumed to be acquired through procedural memory. On the other hand, patients with damaged temporal lobes, which are associated with the declarative memory system, had impairments on tasks where they had to apply irregular past tenses and lexical and vocabulary knowledge (Walenski, Sosta, Cappa, & Ullman, 2009). Corresponding results can be found in people with neurological diseases such as aphasia, and neurodegenerative diseases, such as Alzheimer’s, Parkinson’s or Chorea Huntington (Grossman, Carvell, & Peltzer, 1993; Ullman, 1998; Walenski et al., 2009). For example, Ullman (1998) examined patients with Alzheimer’s disease, which has been linked to damage to the temporal lobe, responsible for declarative memory. When asked to produce the past tense of regular and irregular verbs, Alzheimer’s patients made more errors in irregular verb forms, which are assumed to be stored in declarative memory. The opposite was true for patients with Parkinson’s disease, which is associated with damage to the frontal cortex or the basal ganglia. They made more errors with regular verb forms, therefore showing deficits in their procedural learning (Ullman, 1998).

All in all, this research suggests that declarative and procedural memory both play a role in language learning, but are involved in different language learning processes. This has been shown repeatedly using language tasks in healthy participants as well as in clinical populations with neural damage in areas that are relevant for declarative or procedural memory. To gain further insight in how this relates to the language learning advantage, we will consider the declarative and procedural systems over the lifespan.

(12)

The Two Systems over the Lifespan.

Generally speaking, the acquisition of new skills is thought to peak in adulthood and to then start receding in the late twenties (Craik & Bialystok, 2006). This is also reflected in the maturation of the cortex. Most areas of the cortex do not reach full maturity until early adulthood (Craik & Bialystok, 2006). Research supports this claim for different learning tasks in declarative memory (Craik & Bialystok, 2006). When comparing different age groups, declarative memory was stronger in adult groups compared to children (Anderson & Lajoie, 1996; Kramer, Delis, Kaplan, O'donnell, & Prifitera, 1997; Lum, Kidd, Davis, & Conti-Ramsden, 2010). However, this age-dependent pattern does not appear when looking at the implicit acquisition of skills, a learning process associated with the procedural memory system. When comparing different age groups, researchers failed to find age-specific effects in procedural memory performance across six-year-olds, ten-year-olds and adults (Meulemans, van der Linden, & Perruchet, 1998). Moreover, researchers found that young children aged four to twelve years even show stronger learning effects than older groups (Janacsek, Fiser, & Nemeth, 2012). This finding suggests that children younger than twelve outperform adults on procedural memory tasks.

Children performing on the level as adults in procedural tasks can be explained by the level of maturity of the corresponding brain regions: Regions associated with procedural learning reach maturity in early childhood (Craik & Bialystok, 2006). However, while many studies found that adults and children performed on the same level, some found children outperforming adults. The Declarative/ Procedural Model is not able to explain this without further additions. Hence, the Competitive Systems Approach will be discussed later on.

When applying these findings to language learning, one could argue that aspects of language that rely on declarative memory can be learned just as well by adults as by children. On the other hand, one would expect children to outperform adults in the acquisition of procedural aspects of language, such as phonology or regular grammar. This hypotheses has been tested in previous empirical research (Johnson & Newport, 1989; Sanders et al., 1999). For example, Sanders et al. (1999) showed that late learners of language pick up declarative, lexical information well, but struggle to reach mastery level when it comes to pronunciation, which requires procedural learning processes when acquiring phonetic information.

Interestingly, this suggested age-dependent shift in procedural and declarative memory occurs around the same age that Lenneberg (1967) proposed as critical for first language

(13)

acquisition, namely around the start of puberty. It also corresponds to the sensitive age that has been found in the study by Johnson and Newport (1989) when studying second language acquisition.

In summary, there is strong evidence that there are aspects of language learning that depend on declarative memory, such as semantic learning and irregular grammar (Ullman, 2001). Other aspects of language learning rely on procedural learning, such as phonological learning and grammatical systems (Ullman, 2001). Because of the way declarative and procedural memory develop throughout lifetime (Newport, 1994), one would expect adults to show stronger performance than children in tasks that require declarative learning. This includes aspects of language such as word learning and irregular grammar. At the same time, children should show stronger or equal performance on procedural learning tasks, such as required for phonological learning and grammatical systems. While the maturation of brain regions can explain why adults have an advantage in declarative memory, it cannot clearly explain why children would outperform adults in procedural memory. To explain this, another approach will be discussed.

Competing Memory Systems Approach.

Analogue to the Declarative/ Procedural model, Poldrack and Packard (2003) states that implicit and explicit learning within both systems interact dynamically throughout the life span. More specifically, they hypothesised that the two systems compete for the same neural resources during skill learning. This competition means that when one system is weaker the other system will compensate for it and take over its role in the learning process. A system could be weaker due to lack of maturity or (temporary) damage to the underlying structure. Evidence for this was found in both humans and non-human animals.

Evidence from Non-Human Animals.

In animal research, a compensating nature of the two systems during learning can be observed. Overall, when inhibiting one of the systems, learning in the other system improved (Poldrack & Packard, 2003). Research on this has mostly been performed on rats. Because it is impossible to speak of conscious, explicit knowledge in non-human animals, the memory systems are referred to as hippocampus-dependent (like explicit learning in humans) and non-hippocampus-dependent (like implicit learning in humans). Different kinds of learning can both

(14)

be tested in a water-maze task where the rats are trained to swim to an escape platform (Schroeder, Wingard, & Packard, 2002).

In the hippocampus-dependent place task, the platform is in a fixed place in the maze. All times when a rat takes a corner within the maze are equally reinforced whether they lead towards the platform or not. This means that rats need to “remember” where the platform is, and cannot improve their speed by following the reinforcement. Hence, this kind of learning is similar to explicit learning in humans. In the non-hippocampus-dependent response task, they find the platform at different spots each time. This time, only correct turns get reinforced. Therefore, the rat’s knowledge of where the platform is will not improve their speed. Instead, being sensitive to the reinforcement will lead to improved speed. This is the equivalent of implicit learning in humans.

When injecting a rat’s hippocampus with an anaesthetic drug solution after training, and thereby reducing activation of the hippocampus, researchers found that the performance on the place task decreased while the performance on the response task increased (Schroeder et al., 2002). Inhibiting the hippocampus increases performance on non-hippocampus-dependent tasks and impairs performance on hippocampus-dependent tasks. Similar trends can be observed with rats that have lesions in the hippocampus (McDonald & White, 2013; Packard, Hirsh, & White, 1989). For example, rats that have hippocampal lesions outperform healthy rats in a non-hippocampus-dependent task (McDonald & White, 2013; Packard et al., 1989). This was also true for reversible inactivation of the hippocampus (Schroeder et al., 2002). Other findings concern rats with lesions of the caudate-putamen, a region associated with implicit learning in humans. Those lesions seemed to go along with improved performance on location tasks but impaired performance on response tasks (Mitchell & Hall, 1988).

All in all, there is some evidence for competing memory systems in rats. When one system is impaired, performance on tasks of the other systems are improved. This is the same interaction we would expect to see in human learning experiments and supports the Competing Memory Systems Approach (Poldrack & Packard, 2003).

Evidence from Human Adults.

In humans, similar results have been found. Evidence for this hypothesis often uses the following approach: The PFC is inhibited and performance on an implicit task is measured and

(15)

compared to a control group. Importantly, the PFC is associated with explicit learning (Craik & Bialystok, 2006; Squire, 2004). Thus, we would expect an inhibited PFC to result in stronger performance on implicit tests. To test these effects, different methods have been used. The PFC has been suppressed by using hypnosis (Nemeth, Janacsek, Polner, & Kovacs, 2013), secondary or distraction tasks (Brown & Robertson, 2007; Foerde, Knowlton, & Poldrack, 2006), cognitive fatigue (Borragán, Slama, Destrebecqz, & Peigneux, 2016), alcohol consumption (Virag et al., 2015) or benzodiazepine (Frank, O'Reilly, & Curran, 2006). All these different ways resulted in improved performance in implicit or procedural learning tasks, indicating that a weaker explicit or declarative memory system is associated with a stronger implicit memory system. On the other hand, enhancing cognitive effort by giving corresponding instructions impairs performance on procedural learning tasks (Fletcher et al., 2004; Howard & Howard, 2001), showing that a stronger explicit memory system is associated with a weaker implicit memory system. One should note that all these manipulations target the explicit system directly and only affect the implicit system indirectly. Hence, there is only evidence that a change in explicit memory influences implicit learning. There is no concluding evidence that manipulation of implicit memory would influence explicit memory in the same way.

Furthermore, interactions between the basal ganglia and medial temporal lobe memory systems can be deduced from functional neuroimaging (Poldrack & Rodriguez, 2004). An example of this is the weather prediction task that measures probabilistic classification learning. In this task, subjects predict outcomes (“rain” or “sunshine”) based on a set of cards depicting abstract shapes. Researchers found that during this task, regions in the medial temporal lobe, associated with explicit learning, were deactivated. This deactivation grows more pronounced over the trials (Shulman et al., 1997). The outcome matches the notion that probabilistic learning is mostly implicit. Areas associated with implicit learning, such as the caudate nucleus, showed strong activation during this task (Poldrack et al., 2001), suggesting the competing nature between the two systems. Similar results have also been found for tasks on motor skill learning (Jenkins, Brooks, Nixon, Frackowiak, & Passingham, 1994) and perceptual skill learning (Poldrack & Gabrieli, 2001).

This approach potentially explains why children show superior performance in procedural learning, such as motor or grammar learning, while their cognitive skills are much less developed (Craik & Bialystok, 2006). It may imply that adults are not able to keep up with

(16)

children in some domains of learning because of their superior cognitive skills. Researchers hypothesise that the explicit system is not yet fully developed in children, resulting in weak competition for the implicit system to learn (Poldrack & Packard, 2003). Later in life, however, the explicit system matures and suppresses the implicit system. Consequently, most of the learning done at a young age happens implicitly, while most of the participants of age twelve or higher use explicit techniques. While this puts children at a disadvantage when having to acquire explicit knowledge, such as vocabulary, it grants them an advantage in learning implicit rules underlying basic language structure, such as grammar or phonology.

Competing Systems over the Lifespan.

Multiple recent experiments showed that children have an advantage over adults when implicitly learning different aspects of language, such as grammar or phonology. Researchers have studied speech errors to examine this (Smalle, Muylle, Szmalec, & Duyck, 2017). The basic idea is that even when making speech errors, speakers of a language still adhere to some phonotactic constraints. Phonotactic constraints refer to rules concerning the position of certain phonemes. Smalle, Muylle, et al. (2017) give the example of the phoneme /ŋ/ (e.g. king) which is not allowed to be at the beginning of a word in English. Consequently, it is highly unlikely for a native English speaker to say “ngik” when attempting to say “king” since this kind of speech error would not meet phonotactic constraints. In experimental conditions, this can be used by studying the speech errors participants make when new phonotactic constraints are introduced. When comparing children’s and adults’ speech errors in a four-day experiment, Smalle, Muylle, et al. (2017) found that eight-year-olds implicitly acquire novel phonotactic constraints significantly faster than adults. Adults only managed to learn novel phonotactics after an offline consolidation period, including sleep. This indicates that children are faster implicit learners compared to adults who need more time and exposure to show significant learning.

Another study used the Hebb repetition task, a common paradigm when researching implicit language learning. In this task, sequences of syllables have to be memorised and recalled immediately. One specific sequence (the Hebb sequence) is repeated regularly, with the same amount of trials in between two appearances every time. The recall of that syllable is expected to improve relative to the other sequences (Hebb & Hebb, 1961). Smalle, Page, Duyck, Edwards, and Szmalec (2018) used a version of the Hebb repetition task in their one-year longitudinal study. They used two different Hebb sequences. One was announced, making learning explicit.

(17)

The other Hebb sequence remained un-announced, making learning implicit. In this experiment, children showed the same learning rate as adults despite their lower memory capacity. Furthermore, children were better at retaining the implicit Hebb sequences then adults, both after 4 hours and after 12 months. Lastly, although both groups showed similar performances, adults had more awareness of their knowledge. This indicates that they used more explicit learning strategies (Smalle et al., 2018), as predicted by Poldrack and Packard (2003)

In summary, children seem to outperform adults when it comes to implicit learning (Smalle, Muylle, et al., 2017; Smalle et al., 2018). Moreover, the most prominent learning strategy seems to shift from implicit to explicit when the brain matures (Poldrack & Packard, 2003; Smalle et al., 2018).

Interestingly, this shift from implicit to explicit techniques seems to be reversible. Researchers found that adult learners of second languages can reach not only native-like performances in the language they are learning, but also the same pattern of neural activity (Morgan-Short, Steinhauer, Sanz, & Ullman, 2012). They compared two methods of language learning: implicit language learning, such as via immersion in the language, and explicit language learning, such as via instruction on the underlying structures. Even though both methods allowed learners to reach fluency in the language, only the implicit learning method resulted in native-like neural activity (Morgan-Short et al., 2012). This indicates that implicit learning and explicit learning lead to different neural activity and that adults benefit from implicit studying methods. Moreover, these results point out that adults may still benefit from implicit learning, given they use the appropriate learning techniques.

Similarly, Smalle, Panouilleres, Szmalec, and Möttönen (2017) found that adults who are impaired in their dorsolateral prefrontal cortex (DLPFC), a region associated with explicit learning, via transcranial magnetic stimulation (TMS), perform stronger on the Hebb repetition task. When comparing this to the results from previous studies using Hebb repetition tasks (Smalle et al., 2018), impairment in the DLPFC seems to give an advantage to adults in implicit learning, similar to the one that children have. This seems to further support the notion that adults’ implicit memory does not necessarily worsen with age, but merely gets overshadowed by explicit learning strategies. The explicit, declarative learning system seems to inhibit implicit learning processes that are necessary for language acquisition.

(18)

Overall, these results suggest that implicit, procedural and explicit, declarative learning systems interact competitively during formal aspects of language learning. It seems that explicit learning mechanisms interfere with basal implicit learning mechanisms that are important for language acquisition. Areas that are crucial for explicit learning such as the DLPFC are not fully developed until adolescence (Craik & Bialystok, 2006). This might explain why young children have an advantage in implicit aspects of language learning (namely, less competition from a declarative system). In contrast, adolescents and adults have an advantage in explicit aspects of language learning. Since implicit and explicit strategies seem to result in different neural activity (Morgan-Short et al., 2012) with only implicit learning resulting in native-like proficiency, this has important implications for the ability to acquire language throughout the lifetime.

Taken together, the Declarative/Procedural Model (Ullman, 2001) and the Competitive Systems Approach (Poldrack & Packard, 2003) are able to categorize aspects of language acquisition into declarative and explicit, and procedural and implicit learning. They then establish a compensatory relationship between the two different kinds of learning, with inhibition on one kind resulting in a stronger performance in the other (Borragán et al., 2016; Poldrack & Packard, 2003; Smalle, Muylle, et al., 2017). Because the implicit learning systems matures earlier than the explicit one (Craik & Bialystok, 2006), children’s explicit learning is assumed to be weaker, and therefore the implicit learning compensates and performs stronger than it would in adolescents or adults. This provides a potential explanation for the language learning advantage in children. In order to test this, a language-related paradigm with a focus on implicit learning is required. In the following chapter, such a paradigm will be depicted, alongside previous findings and recent variations of it.

Statistical Learning

In order to describe the statistical learning (SL) paradigm, we first need to establish the term SL and its relevance for language learning. Similar to implicit learning, SL is a process where individuals acquire knowledge without intent or conscious effort. SL is described as the process of becoming sensitive to statistical structure in the environment (Batterink, 2017; Saffran, Aslin, & Newport, 1996; Saffran, Newport, et al., 1996). This process is based on subconsciously noticing rules and patterns in the world around us and adjusting our expectations accordingly. After some time, we have acquired enough information from the environment to predict what

(19)

will happen next according to the pattern. SL can happen for all modalities, and therefore does not apply exclusively to language learning.

SL is described as being involuntary and automatic (Fiser & Aslin, 2001, 2002), happens without awareness (Turk-Browne, Jungé, & Scholl, 2005), and as a by-product of repeating exposure (Saffran, Johnson, Aslin, & Newport, 1999). Essential features of SL are that it occurs without instructions or conscious attempts, such as when stimuli are presented without any task (Fiser & Aslin, 2001, 2002; Toro, Sinnett, & Soto-Faraco, 2005) or with an unrelated cover task (Toro et al., 2005; Turk-Browne et al., 2005; Turk-Browne & Scholl, 2009). As such, SL is closely related to implicit and procedural learning (Batterink, Paller, & Reber, 2019). However, the processes of SL can also result in explicit, declarative knowledge (Perruchet, Gallego, & Savy, 1990; Perruchet & Pacteau, 1990; Servan-Schreiber & Anderson, 1990). This result happens when participants become not only sensitive to but aware of statistical structures in the tasks.

Interestingly, recent studies demonstrated that SL skills could vary depending on modality and stimulus and does not seem to depend strongly on general cognitive abilities like intelligence or working memory (Frost, Armstrong, Siegelman, & Christiansen, 2015; Siegelman & Frost, 2015). Though research on SL often focuses on auditory stimuli (e.g. speech segmentation), there have also been studies using visual stimuli (Fiser & Aslin, 2002; Turk-Browne et al., 2005) and tactile stimuli (Conway & Christiansen, 2005). In the following, the specific influence of SL on language acquisition will be elaborated upon.

SL and Language.

SL is of relevance on multiple levels of language learning. Its close connection to procedural learning implies that it would be of relevance to similar aspects of language, such as regular grammar (Kidd & Kirjavainen, 2011; Lum & Kidd, 2012; Ullman, 2001), syntax (Saffran, 2002) and phonology (Ullman, 2001). One of the most commonly researched aspects of language that includes SL is speech segmentation. This describes learning how to separate words from each other in a stream of sentences. Speech segmentation is a process anyone goes through while acquiring a new language. First, spoken speech in an unknown language seems like a continuous stream of sounds without any reliable pauses between words (Lehiste, 1960). Then, through SL, the boundaries between separate words are learned. Saffran (2003) used the phrase “pretty baby” to demonstrate this. When exposed to the English language, we are quite likely to hear the syllable “pret” followed by the syllable “ty” because they are in the same word.

(20)

When hearing this word repeatedly, we will make this association and understand that those syllables belong together and form a word. Even though in the phrase “pretty baby”, the syllable “ty” is followed by the syllable ”ba”, we are unlikely to hear this combination of syllables a lot when exposed to language. After a certain period of time, we will deduce that “ty” and “ba” do not belong together as a word, since they appear more often separate from each other than together. Without acquiring this statistical knowledge, it would be next to impossible to learn to understand any spoken language since words cannot be separated from each other. The ability to compute words this way has been found in infants as young as eight months (Saffran, 2001; Saffran, Aslin, et al., 1996).

SL and its Components.

SL can be divided into two components (e.g., Batterink & Paller, 2017). Firstly, the word

identification component describes how an individual’s starts to perceive and encode individual

units, such as syllables or phonemes, as integrated items, such as words. It poses the central challenge of SL. This challenge comes down to finding the right chunk and establishing it as a word. Secondly, the memory storage component influences performance on tests of SL (Batterink, 2017). It influences the storage of representations in long-term memory. It is, therefore, a prerequisite for further processing, such as acquiring phonological patterns across words (Saffran & Thiessen, 2003) and mapping words to objects (Estes, Evans, Alibali, & Saffran, 2007; Mirman, Magnuson, Estes, & Dixon, 2008). Most importantly, memory storage strongly influences the results of all measures that occur after learning is completed, as opposed to during learning. Thus, the way in which we measure SL will influence to what extent which component is measured.

As a result, it can be challenging to separate the two components within experimental research. Traditionally, participants often acquire knowledge during a period of exposure, and are then asked to remember it in the task. This requires them to use both components and offers no possibility to separate them. In the following, an often-used auditory SL paradigm will be described, as well as recent variations that allow us to gain further insight into the specific kind of learning that has occurred during exposure – for example by separating the SL components or by dividing the measure into explicit and implicit knowledge.

(21)

The SL Paradigm and Recent Variations.

One of the most commonly used experiments to research SL in language learning is the artificial speech segmentation paradigm (Saffran, Aslin, et al., 1996; Saffran, Newport, et al., 1996; Saffran, Newport, Aslin, Tunick, & Barrueco, 1997). In this task, participants are first exposed to a continuous stream made up of repeating trisyllabic nonsense-words (e.g. bupada + babupu + tutibu + …,). However, there is no a priori knowledge or cue for where the word boundaries are. The only cue available is statistical in nature: Syllables that belong to the same word are more likely to occur together than syllables from two different words, just like in Saffran’s “pretty baby”-example (Saffran, Aslin, et al., 1996). To measure whether participants indeed detect word boundaries and acquire new words based on SL, they are asked to discriminate words from non-words after the exposure. Non-words are further trisyllabic nonsense-words that did not occur during exposure. Thus, participants should be able to discriminate the words from the exposure from the non-words if they learned, whether implicit or explicit, about the underlying statistical cues. This learning process is tested by presenting one word and one non-word at a time and asking the participant to choose which one they recognise. The idea is that when the overall accuracy surpasses chance level (50%), participants have obtained word knowledge via SL processes. With this task, also called a forced choice

recognition task, numerous studies showed that both children and adults can achieve statistical

language learning (Batterink & Paller, 2017; Batterink, Reber, Neville, & Paller, 2015). The task is considered an indirect, offline measure – meaning that it is obtained after learning has already occurred. Because of the key characteristics of SL, one would expect it to mostly result in implicit knowledge (Conway & Christiansen, 2005; Turk-Browne et al., 2005). The forced choice recognition task is, however, not able to discriminate between implicit and explicit knowledge. It has been found to also reflect explicit knowledge in adults (Perruchet et al., 1990; Perruchet & Pacteau, 1990; Servan-Schreiber & Anderson, 1990).

Recent research suggested an addition to the task that might be able to separate the two kinds of learning: participants are asked to report their level of confidence in their answer (Batterink et al., 2015). This approach is based on the notion that knowledge is implicit when participants do not have meta-knowledge of (i.e. are not aware of) what they have learned (Dienes & Berry, 1997). There are two possible criteria to determine whether this is the case: The guessing criterion and the zero-correlation criterion. The guessing criterion is fulfilled as

(22)

soon as the participant is not aware of the knowledge and assumes to be just guessing, but still performs above chance level. The zero-correlation criterion is fulfilled when the performance of a participant and the confidence in their performance do not correlate (Dienes & Berry, 1997). These criteria can be used for tasks where the ability of a participant to recognise an item from previous exposure is tested, such as the forced choice recognition task (Dienes & Scott, 2005). In other words, participants have to not only detect the correct word but also indicate why they recognised it: do they remember, does it seem familiar or are they just guessing? The remembered words should represent explicit knowledge, while the guessed words should represent implicit knowledge. Using this addition, previous researchers confirmed that mainly explicit knowledge was acquired through SL in adults (Batterink & Paller, 2017; Batterink et al., 2015): they lacked meta-knowledge of the words as indicated by chance behaviour on guessed trials. Since implicit and explicit learning seem to interact dynamically throughout the lifespan, measuring them separately seems crucial for understanding the sensitive age-hypothesis in language learning. Special attention should go to the comparison of children before and after the sensitive age of language learning, as well as adults on this type of memory judgement procedure.

While the addition of a confidence rating allows researchers to differentiate between implicit and explicit learning, all data is still collected after the exposure, and therefore after learning has already occurred. As mentioned before, this can lead to difficulties when trying to separate the different components of SL, namely the word identification component and the memory storage component. This inability possibly leads to an underestimation of SL due to the decrease of knowledge between exposure and recall (Batterink & Paller, 2017). Because children’s working memory is less developed (Newport, 1994), it seems likely that this delay could cause an underestimation of the critical age effect for language learning. While children might have learned more during the actual exposure, they might have had a harder time recalling or reflecting upon this information later on. At the same time, adults might have learned less, but their recall or reflection on it is better. The difference in overall performance between the two groups might then be underestimated, or not found at all. In order to circumvent this problem, learning would have to be also measured while it is happening. This would be considered an online measure of SL.

(23)

Recent research has introduced possible online measures, such as reaction time (RT) to a target syllable, LPC amplitudes, EEG-based measures. (Batterink & Paller, 2017; Batterink et al., 2015). Online measures are taken during the exposure, not after. They assume that each trisyllabic word (e.g. “bupada”) has an unpredictable syllable, namely the initial one (“bu”) and two predictable syllables, namely the medial (“pa”) and final (“da”) syllables. Throughout the exposure, participants should acquire the ability to predict medial and final syllables, knowing that an initial syllable (“bu”) is always followed by the medial and final syllables (“pada”). However, they should not be able to predict the initial syllable, since it always follows the final syllable of a different word (e.g. “debu”, “habu”, “efbu”, …). Importantly, this statistical knowledge about possible predictions usually stays subconscious and implicit. When presented with a target stimulus before exposure (target detection task), we would expect participants to react to all three syllables in approximately the same way at the beginning of exposure since they have no statistical information on any of the syllables yet. At this stage, participants cannot predict any of the syllables. However, with an increasing amount of exposure, we would expect participants to show higher levels of statistical knowledge for medial and final syllables than for initial syllables. For RT, this means that we do not expect participants to increase RT for all syllables. Instead, participants would become increasingly faster when reacting to the predictable medial (“pa”) and final (“da”) syllables, but not when reacting to the unpredictable initial (“bu”) syllables. The difference in RT between the mean of the predictable (medial and final) and unpredictable (initial) syllable then serves as an online measure of SL (Batterink & Paller, 2017; Batterink et al., 2015).

During the exposure, LPC amplitudes, more specifically P300, can be measured. P300 describes an event-related potential component that can be detected as a reaction to a target stimulus (Polich, 2007). It can be observed when participants are asked to discriminate a target stimulus from a stream of sounds, with its amplitude negatively correlated to target probability (Duncan-Johnson & Donchin, 1982; Polich, 2007). For the word segmentation paradigm, this indicates that P300 should be high for all target syllables at the start of exposure, and reduce for the predictable medial and final syllables over the course of the experiment, while it should remain high for the unpredictable initial syllable. This was confirmed in a recent experiment (Batterink, 2015).

(24)

EEG-based measures can measure SL by taking advantage of the neural steady-state response. This response resonates at the same frequency as an ongoing rhythmic stimulus (Buiatti, Peña, & Dehaene-Lambertz, 2009), such as the words or syllables in the target detection task. The idea is that the steady-state response at the frequency of the syllables will be strong at the start, while it will be weak at the frequency of the word. In other words: participants will perceive individual syllables and not trisyllabic words. The hypothesis is then that successful SL would result in a decrease at the frequency of the individual syllables and an increase in the frequency of the trisyllabic words. Batterink and Paller (2017) confirmed this in a recent experimental study.

Some recent evidence indicates that online measures might be more reliable indicators of SL than offline measures. This stems partially from the notion that the offline measure does not only reflect the total acquired knowledge but also how well the participants have remembered and are able to recognise them (Batterink et al., 2015). Recent studies suggest that this could lead to offline measures underestimating SL (Batterink et al., 2015). Moreover, while offline measures reflect both explicit and implicit knowledge, online measures are assumed to reflect exclusively implicit learning (Batterink et al., 2015). The relationship between the online and offline measures of SL needs to be further investigated. While some research suggests a positive relationship (Batterink & Paller, 2017), other studies have found that they are unrelated (Batterink et al., 2015).

SL over the Lifespan.

Even though the SL paradigm is inherently linked to language acquisition, little research explored SL over the course of the human lifespan. While Saffran, Newport, et al. (1996) found that infants learn within two minutes of exposure, older children and adults seem to need longer exposure to achieve significant learning on the recognition task in a similar experiment (Saffran, Aslin, et al., 1996). However, as far as we know, no direct comparison between children and adults has ever been done. Furthermore, more recent research has found adults to gain some level of statistical insight after as little as one trial (Batterink & Paller, 2017).

However, when dividing offline SL into implicit and explicit learning using a memory judgement task, Batterink et al. found that adults mostly acquire explicit knowledge, and no significant amounts of implicit knowledge when compared to chance level (Batterink et al., 2015). Since much evidence points to children outperforming adults when it comes to implicit

(25)

learning (Janacsek et al., 2012; Nemeth, Janacsek, & Fiser, 2013), this leads to the question whether children would be able to outperform adults when it comes to SL. This seems especially likely for the “guess”-trials of the forced choice recognition task and the online target detection task, which are associated with implicit learning.

Present Study

The present study aims to investigate the sensitive age period for language learning using a SL approach. Specifically, it aims to take the first steps towards answering two questions: Are children better at acquiring new languages, and if so, why is that the case? Young children and adults were tested using recent variations of the artificial speech segmentation paradigm (Saffran, Aslin, et al., 1996; Saffran, Newport, et al., 1996; Saffran et al., 1997) which allowed us to discriminate between implicit and explicit learning, and between online learning and offline memory storage.

Specifically, we used the offline forced choice recognition task with two recent variations. Firstly, we added a confidence rating to the offline measure (Batterink et al., 2015). These variations allowed us to discriminate between implicit and explicit learning by asking the participant whether they either knew the answer (i.e., they had explicit memory for it), whether they had no clear memory for it (it seemed familiar) or whether they were guessing. Past research suggests that performance above chance level only reflects implicit learning for trials where participants report to be guessing, but not for those where they claim to remember the right answer (Batterink et al., 2015; Dienes & Berry, 1997). Discriminating between implicit and explicit learning is of crucial importance to research that compares children and adults since these learning systems interact dynamically throughout the lifetime according to the competing system hypothesis (Poldrack & Packard, 2003).

Secondly, we added an online measure based on RT to a target stimulus (Batterink & Paller, 2017; Siegelman, Bogaerts, Kronenfeld, & Frost, 2018) which allowed us to measure learning while it is occurring. This was done by measuring the time that the participants needed to detect a specific syllable, and then calculating the difference in mean RT to predictable syllables (medial and final) syllables, and unpredictable syllables (initial) syllables. The difference should reflect the amount of statistical knowledge that the participants have acquired. Recent evidence suggests that online measures might be a more reliable measure of implicit learning (Batterink & Paller, 2017). Moreover, it did not only let us investigate differences in the final

(26)

level of statistical knowledge, but also in the learning process of the different groups. This distinction is crucial since children show different learning patterns than adults in other processes that rely on procedural memory (Craik & Bialystok, 2006), and research on SL has found contradictory results when only looking at offline measures (Batterink & Paller, 2017; Saffran, Newport, et al., 1996).

As the last addition, we used the Dutch version of the Peabody Picture Vocabulary Test (PPVT, Dunn, Dunn, & Schlichting, 2005) in order to test the participant’s language skill. This allows us further insight into the role of SL in applied language skills and account for the ecological validity of our results.

To summarise, groups of children and adults completed an online target detection task while being exposed to continuous speech streams. Afterwards, they completed the PPVT. Lastly, they were asked to complete the offline first choice recognition task with the addition of a confidence rating.

Hypotheses.

Based on the Declarative/ Procedural Model (Ullman, 2001) and the Competing Memory Systems hypothesis (Poldrack & Packard, 2003), we expected younger children to outperform adults in the overall measures of SL. This age-effect was assumed to be most visible in the implicit measures of SL, such as the online measurement of RT and the implicit part of the offline measurement, but not in the explicit part of the offline measurement.

For the forced choice recognition task, we expected both groups to show significant learning. As the memory judgement (i.e. remember, familiar, guess) was supposed to separate implicit from explicit learning, we expected the different judgements to influence the accuracy. We expected this to happen differently for children and adults. We expected stronger performance from adults when it came to explicit learning and stronger performance from children when it came to implicit learning. Specifically, we expected the following for the forced choice recognition task:

1. Children were expected to perform above chance when claiming to guess, while adults were expected to perform on chance at guessed trials. This hypothesis would indicate that only children have formed implicit memory for the novel words (while adults have mainly formed explicit memory).

(27)

2. Accuracy on guess trials was expected to be higher in children than in adults. For remembered trials, we expected better performances from adults than from children.

For the online target detection task, we expected corresponding results. Since this measure mostly reflects implicit learning, we expected children to outperform adult on this measure.

3. Both children and adults are expected to achieve significant SL effects in the online target detection task. However, we expected there to be a significant effect of the group, indicating that children implicitly learn better than adults.

4. Children were also expected to learn faster than adults. Learning trajectory would be measured by dividing the exposure stream into separate trials blocks.

Lastly, we investigated the relationships between the different measures of language learning to indicate internal and external validity of our laboratory measures. Specifically, we expected the following:

5. The different measures of language learning, namely the online measure, the offline measure and the norm score of the PPVT, correlate positively.

(28)

Method

Participants

In total, a sample of 102 participants was tested for this study. Eight participants had to be excluded for data analysis. Four of those were excluded because they were diagnosed with a developmental disorder, one because they did not speak Dutch fluently, and three more because of technical failures during testing. The remaining sample consisted of 47 adults (10 males) and 47 children (15 males), who were aged between 18 and 26 years (mean: 23.08 years,

SD: 2.47 years), and between 7 and 11 years (mean: 9.00 years, SD: 1.05 years), respectively.

The adults were recruited via social media and a university internal recruiting system which grants students credits for participation. The children were recruited by contacting local schools and organisations, as well as via social media. Before the experiment took place, an informed consent was signed. For the children, a parent of the participant signed an informed consent. Education level was asked in a short questionnaire as a measure of socio-economic status of the participant. For children, the education level of the mothers was asked. The data is summarised in Table 1.

Table 1

Summary of the education level of participants or their mothers (for children).

Lower education1 _{High school}2 Bachelor’s degree3 Master’s degree

or higher4

Children 7.5% (3) 15% (6) 47.5% (19) 30% (12)

Adults 0.0% (0) 34% (16) 42.6% (20) 23.4% (11)

1 _{Vocational, artistic, technical or general lower diploma.}2_{Vocational, artistic, technical or}

general high school diploma (12 years of school), 3 _{academic or professional}4 _{one or more}

(29)

Materials

A Dell Latitude E5520 laptop with a 15.6inch monitor (1366 x 768; 60Hz) and Windows 7 (32 bit) was used to administer the data.

For the SL paradigm, we selected letters of the alphabet that result in a consonant-vowel or vowel-consonant structure when produced (e.g. “Be”, “We”, “eF”). We chose to use stimuli from the Dutch alphabet as they sound familiar for young children. Moreover, it ensures that the novel to-be-learned words contain sounds that exist in spoken Dutch. In total, 12 syllables were randomly selected from the pool of alphabet letters, and recorded by a female native Dutch speaker in neutral intonation. They were then edited using Audacity software (Team, 2008) so that they had a duration of 300ms each. Four trisyllabic pseudowords (henceforth referred to as words) were created using those syllables. Another four trisyllabic pseudowords (henceforth referred to as non-words) were created using the same syllables in a novel order (see Figure 1 for an overview of the randomisation procedure).

For the online target detection task, continuous speech streams were created by combining the previously created words in a pre-set order. Note that the streams were made by combining the individual audio files of the syllables without any breaks. Hence, there were no

eLQuPe

eMeRDe

VeGeeS

CeKaeF

VeGeeS

eMeRDe

CeKaeF

eLQuPe

Continuous Speech Stream

Figure 1. Overview of the Randomization Procedure for the trisyllabic pseudowords used in

the SL tasks.

Inventory of letters from the

Dutch alphabet: Be Ce De eF Ge Ha Je Ka eL eM Pe Qu eR eS Te Ve We 12 randomly selected: Ce De eF Ge Ka eL eM Pe Qu eS Ve We

Four trisyllabic words:

eLQuPe eMeRDe VeGeeS

CeKaeF Four trisyllabic non-words:

VeKaeL

CeGeeM QuDeeF

(30)

cues for when the words begin or end, except for statistical ones. The order of the words differed in three conditions, with the restriction that the same word would never occur consecutively.

The task consisted of four blocks separated by three breaks. Each block consisted of multiple streams. In each stream, there was one target syllable that participants would be asked to detect. Each syllable of the four words functioned as a target one time per block, resulting in 12 streams per block and 48 streams total. Hence, target position, (i.e., the position of the target syllable within the word) was automatically balanced over the different streams. Each word occurred six times per stream. This results in a total stream length of 72 syllables or 21.6 seconds. Furthermore, each word was repeated 288 times (i.e., 72 repetitions per block and 6 repetitions per stream). The online detection task was performed using Presentation® software (Version 18.1, Neurobehavioral Systems, Inc., Berkeley, CA, www.neurobs.com).

In the offline forced choice recognition task, participants listened to one word and one non-word per trial. They would only hear them once in a pre-set order. They were then asked to recognise which of the two they had already heard during the online task. Afterwards, they were asked about the confidence of their reply: did they remember the word, did the word sound familiar but without clear memory for it, or did they guess? Each participant compared each word with each non-word, resulting in a total of 16 comparisons. Two versions of this task were used, differing only in the order of presentation of word and non-word. The offline forced-choice target detection task used Psychopy 3 software (Peirce et al., 2019).

The Dutch version of the PPVT (Dunn et al., 2005) was used as a measure of receptive vocabulary. It contains a total of 19 sets of 12 items. For each item, four pictures are shown and the examiner reads out a word. The participant then has to choose a corresponding picture (e.g. if the examiner reads “ball” the participant has to indicate the picture with a ball on it). Each participant starts at a set that corresponds to their age. If they make more than one mistake in the starting set, they will move back in difficulty until they can complete a set with one or less mistake. They then move forwards in difficulty until they make more than eight mistakes in a set. Afterwards, a norm score is determined, allowing comparison with a group of the same age. Procedure

The experiment took place in a quiet room. Before the experiment began, participants gave their informed consent and filled in a general questionnaire, including questions about age,

Do children outperform adults in implicit language acquisition? Evidence from a statistical learning paradigm