KIEL CHRISTIANSON AND SARAH‐ELIZABETH DESHAIES
Prior to a certain age – let us say the age when formal education begins – all language learning is “informal.” Any discussion of the cognitive underpinnings of informal language learning must therefore include at least some basic considerations of first‐language acquisition. Beyond that, though, it must also wrestle with differences between child and adult second‐language acquisition, and the persistent problem of why children appear to be essentially uniformly adept at informal language acquisition, whereas adult performance ranges from adept to abysmal. Put another way, a discussion of language learning in informal environments needs to address how humans acquire language throughout the lifespan, as well as how the ability to acquire language might change across the lifespan.
Charles Yang, in his book The Price of Linguistic Productivity: How children learn to break the rules of language (2016), sums up language learning thus: “I envision language learning as a search for productive generalizations” (9). One thing necessary, of course, for the success of this search is the ability to identify productive rules, like “add –ed to make English verbs past tense.” There are yet two more necessities, namely the ability to recognize both non‐productivity and, more importantly, to recognize exceptions to or gaps in productivity. In English first language acquisition, children tend to follow a U‐shaped learning curve as they acquire morphology (Marcus et al. 1992; Pinker 1999), such that early exceptions in morphological paradigms are observed in children's production, like run‐ran, before being overregularized for a while (run‐runned), before eventually returning to their exceptional forms. As another example, consider the following anecdote involving the first author's son at the age of about 3.5 years: one day, his preschool teacher reported at pick‐up that he had said “the funniest thing” at school. When he had been given a sort of frilly shirt to wear during dress‐up, he had asked for something “more boyly.” The teacher wondered whether that was a word the family used at home. It was not, which surprised the teacher. This was not a case of “hear and repeat”; rather, he had simply extended a faulty paradigm – manly – womanly – girly – *boyly – to include the fourth form, which, rather mysteriously, does not exist in Standard American English. Sadly, this delightful overregularization disappeared from the child's language as quickly as it appeared (and now at 15, he has no recollection of it). That errors of this sort should be so transient, however, is to be expected if language learners – or at least child‐language learners – have access to some mechanism that allows them to sort “rules” from “exceptions.” Errors made by child‐language learners do not “fossilize,” as many adult second language (L2) learners do (Selinker 1972).
What is the mechanism, then, that allows children to apply rules productively, filter out exceptions, and recognize holes, or “leaks,” in grammars (cf. Sapir 1928) and adjust their linguistic behaviors accordingly? That, of course, is the billion‐dollar question. If we knew the answer, adults could optimize or hone that mechanism to operate over linguistic input in informal environments (and formal environments) to learn languages just like, and just as efficiently as, children do. Most broadly, the candidate mechanisms can be classified as domain‐specific or domain‐general. Cognitive researchers refer to a cognitive process or mechanism as being domain‐specific if it is seen as having a specialized or segmented function within the brain (see Fodor 1983). Conversely, if a process or mechanism is viewed as domain‐general, it is considered to be emergent from multiple systems operating over the input, such as memory, attention, speech perception, and pattern recognition (Saffran and Thiessen 2007; Tomasello 2009). Considering language acquisition – either first language or second language acquisition – solely within this broad dichotomy can lead to confusion, however. We must also establish whether we are discussing language knowledge (i.e. grammar or language form) or the mechanisms by which linguistic input is analyzed as it is used to determine the shape of the end‐state knowledge. Losing sight of this second dichotomy opens the door to severe misunderstanding. The domain‐general learning mechanism of statistical learning (SL; e.g. Frost et al. 2015) is often marshaled, incorrectly, as evidence against the existence of domain‐specific linguistic knowledge, i.e. of an innate language endowment (cf. Chomsky 1957, 1959, 1975, 2000) that is in some sense modular (cf. Fodor 1983, 2001) within the human mind. As Jerry Fodor pointed out explicitly, “Chomsky can with perfect coherence claim that innate, domain specific PAs [propositional attitudes, i.e. knowledge structures] mediate language acquisition, while remaining entirely agnostic about the domain specificity of language acquisition mechanisms” (2001, 107–108). Chomsky himself (1955/1975, as cited by Yang 2008) explicitly supported the exploration of probabilistic language learning mechanisms. For example, he states, “… [It] might turn out to be the case that statistical considerations are relevant to establishing, e.g. the absolute, non‐statistical distinction between [grammatical vs. ungrammatical]…. Note that there is no question being raised here as to the legitimacy of a probabilistic study of language” (1955/1975, Section 36.3).
Given the above discussion and clarifications, then, we are left with the perennial puzzle: What is it that children do to be such efficient language learners that adults seem to not be able to do? Furthermore, we are left with two distinct sources of potential deficiencies in adult L2 learning: knowledge or mechanism. From the 1970s into the 1990s, numerous investigations were conducted into whether or not innate linguistic structure, dubbed “Universal Grammar” (UG) by Chomsky, was accessible to adult L2 learners. Perhaps the most influential exploration of this topic was Lydia White's book on the topic, Universal Grammar and Second Language Acquisition (1989; see also White 2003). These investigations were inspired in large part not only by Chomsky's work, but also by Eric Lenneberg's book, The Biological Foundations of Language (1967), in which a “critical period for language acquisition” was proposed. Yet White, like Chomsky before her and Fodor after her, acknowledged that domain‐general, probabilistic mechanisms such as SL could exist perfectly coherently with innate linguistic knowledge structures. As such, it is possible that empirical results suggesting that adult L2 learners do not have access to UG might instead be attributed to misapplications of statistical regularities, or apparent regularities, in the L2 input. In an excellent review of the critical period hypothesis, Hyltenstam and Abrahamsson (2000) conclude that there is more evidence for the existence of maturational constraints on “native‐like” second language learning than there is evidence against these constraints. Still, we are left to wonder what the hypothesized constraints constrain (see a recent empirical study into this question by Nowack and Baggio 2017). There appears to us to be a paucity of compelling evidence that the human cognitive architecture within which linguistic knowledge is situated differs between children and adults. On the contrary, adults would seem to have several significant cognitive advantages over children in learning any topic in informal environments, including language. We will therefore assume for the remainder of this chapter that what distinguishes child L1 learners from adult L2 learners in informal language‐learning contexts is their respective abilities to extract statistical regularities from the ambient language input and to apply those regularities to their respective L1 and L2 grammars (cf. Hudson Kam and Newport 2005, 2009; Newport 1990; Scovel 2000). As a concrete example, we point to the results of Granena and Long (2012), who found negative correlations between age and rate of learning of L2 phonology, lexis, and morphosyntax, taking into account length of residence and language aptitude. Specifically, declines were observed first for L2 phonology (age correlations obtained in the 3–6 year‐old range), then lexis (correlations obtained in the 7–15 year‐old range), and finally morphosyntax (correlations obtained in the 16–29 year‐old range, in particular a large drop‐off in the late teens).
Yang (2002, 2004) proposes the variational approach to L1 language acquisition, in which children use domain‐general mechanisms to discover generalizations about their particular language within the search space demarcated by UG. Situating domain‐general learning mechanisms, specifically SL, within larger domain‐specific cognitive “spaces” accords well with some recent SL literature. Frost et al. (2015) propose a similar architecture to help account for observations that SL fails to generalize across cognitive domains or even across modalities within a domain. Erickson et al. (2016) similarly report that various measurements of SL generally fail to correlate either across, or even within, individuals. This result suggests that although SL might well be a generalized learning mechanism (for language and for other things), the efficacy with which it is applied might well vary considerably from person to person and from domain to domain. So if we at least for now assume the “search space” delineated by UG is the same for adults and children, but that adults might differ with respect to how and how well they navigate that search space, what must adult L2 learners in informal language environments be able to do to learn the L2 with some semblance of proficiency?
The first, and most salient, feature of informal language learning is that unlike formal instruction, the input in informal environments is not structured. There are no neatly printed tables of inflectional paradigms or tenses, no lists of “word families,” or pictures of scenes with the L2 labels scattered helpfully throughout. Given that the input is unstructured, the adult L2 learner, like the child L1 learner, would need to apply SL mechanisms to the flood of input to separate the productive “wheat” from the exceptional “chaff.” The first thing that the adult language learner would need to do, then, is to figure out where to look in the input for information that would allow them to start making structural generalizations. One might speculate that adults could apply sophisticated metalinguistic strategies to the input to help with this task. And indeed, it has been observed repeatedly in the literature that adult language learners with greater metalinguistic awareness – essentially learners who are predisposed to thinking about language as an object of study, which is fundamentally systematic and rule‐governed with inherent regularities – are more successful learners. Research examining the metalinguistic awareness of language learners is increasingly being carried out in multilingual populations, where it is being shown consistently that the more linguistic experience one has, the more metalinguistic awareness one has. For example, Ransdell et al. (2006) found that not only are bilinguals and multilinguals more accurate than monolinguals in self‐assessments of their own linguistic capacities, they also scored significantly better on a reading span task measuring verbal working memory (see also Bialystok 2007, 2010).
What, concretely, does metalinguistic awareness buy adult L2 learners in informal learning environments, though? For one thing, it allows the learner to divide incoming L2 input into categories over which SL algorithms can operate. Consider a hypothetical learner of L2 Spanish (or a not‐so‐hypothetical learner, such as the first author's son again) who is not able to “feel” or “see” the distinction between ser (usually used for permanent, unchangeable attributes) and estar (usually used for temporary, changeable attributes). Although both the metalinguistically aware and metalinguistically unaware learner will be faced at first with seemingly arbitrary pairings of these copular verbs and various attributes, only the former will realize that there must be some systematicity to the alternations and work to determine what that system is. The same goes for English separable verbs and verb + preposition constructions such as the following: look up a word/look a word up vs. climb up a ladder/*climb a ladder up; run out the clock/run the clock out vs. run out the door/*run the door out. To the metalinguistically unaware learner, pairs such as these look hopelessly confusing and contradictory. To the metalinguistically aware learner, they hint at some underlying rule‐governed system, and the structures in question must differ qualitatively in some important way. The former learner may throw up his or her hands; the latter will likely seek out more examples and work, consciously or unconsciously, to discover the relevant characteristics of the different verb types and the rules governing them.
Closely related to metalinguistic awareness is being able to consciously apply learning strategies to language learning. This is especially important in informal learning environments, as the L2 input is encountered in an unstructured manner. Conscious application of strategies would also seem to be something that adults could do better than children; however, it does not appear to be something that comes automatically to adults. Studies have shown that language learners with more language experience (e.g. trilinguals and multilinguals vs. monolinguals and bilinguals) are more efficient and successful applying learning and study strategies in the learning of a new language (Cenoz and Todeva 2009; Kemp 2007; Psaltou‐Joycey and Kantaridou 2009). Most of this research has been done in the area of third language (L3) acquisition (e.g. Cenoz 2013), though, so much more work is needed to understand how adult L2 learners with variable success in L2 learning in informal contexts may vary in their metalinguistic awareness and learning strategies.
Explicit learning strategies can be contrasted with implicit learning, usually associated (again) with SL. Nick Ellis (2002, 2005) argues that most first language learning is implicit, resulting in tacit knowledge of all levels of linguistic analysis from phonology through pragmatics via distributional analysis of the input. L2 learning, however, is not nearly as amenable to implicit learning, especially by adults. In particular, errors are more apparent to adults, and trigger activation of explicit monitoring and applications and revision of rules. This error‐driven learning process is described by Ellis (2005) as an interface between implicit learning and explicit learning. An error is hypothesized to be a key event triggering the conscious noticing (Schmidt 1984) of some aspect of the input, generating an explicit memory of some aspect of the input that then gets integrated and consolidated into implicit memory.
In an early comparison of implicit vs. explicit L2 construction learning, Shin and Christianson (2012) compared three structural, or syntactic, priming conditions: pure long‐lag structural priming, no‐lag structural priming, and no‐lag structural priming with explicit instruction. The term priming refers broadly to an effect whereby exposure to a given stimulus A facilitates the processing of or memory for a subsequent stimulus B. With respect to structural priming in language, Bock (1982; see also Bock and Griffin 2000; Bock and Loebell 1990) first showed that people exposed to passive sentences or dative (indirect object) sentences are more likely to subsequently produce passives or datives than people who have not been exposed to these structures first. In other words, if you hear a dative sentences like The postal worker handed the doctor a letter, your chances of describing a picture of a woman playing catch with a child as The woman threw the child a ball are increased compared to a no‐prime baseline condition (in which nearly all descriptions will be the more frequent prepositional object, The woman threw the ball to the child). These sorts of results have been attributed to both implicit learning (cf. Chang et al. 2000, 2006) and explicit lexical activation account (cf. Pickering and Branigan 1998). Shin and Christianson (2012) asked whether L2 learning of structures like the dative construction and separable participle verbs (e.g. put out the fire vs. put the fire out) would differ between implicit priming (long‐lag structural priming, in which several filler sentences intervened between stimulus A and stimulus B), more explicit priming (no‐lag priming, drawing conscious attention to stimulus A and stimulus B), and no‐lag priming paired with explicit instruction (e.g. followed by a schematic word order note). Furthermore, Shin and Christianson brought all participants in for two sessions: the original learning session (which included pretests and posttests on the structures) and a second session the following day (which included a second, delayed posttest) in order to assess the effect of intervening sleep and memory consolidation on language learning. Participants were college‐aged native speakers of Korean (N = 48) who were L2 English speakers.
The results are, very briefly, presented in Figure 2.1. In short, explicitly reinforced instruction facilitated immediate learning of both the dative constructions and the phrasal‐verb structures. In the case of the dative constructions, the facilitative effects of both the no‐lag priming (i.e. explicit priming) and explicitly reinforced priming conditions fell off at the second post‐test (24 hours later). So although the purely implicit long‐lag priming was not as facilitative as the explicitly reinforced priming in the immediate posttest, it was just as good as the explicitly reinforced priming in maintaining learning the next day. The story with the phrasal verbs was different, with all three conditions performing equally well in immediate posttesting, but the explicitly reinforced condition maintaining its advantage the following day. This pattern was explained with reference to recent work examining the consolidation of procedural memory during sleep. It was argued that syntactic structure‐building is a procedural, implicit task, whereas memorizing and recalling phrasal verbs is more of an explicit task. Whether this interpretation is accurate remains open to further experimentation; however, the results reinforce Ellis's (2005) central theme that implicit and explicit learning and memory must interface, and strongly suggest that different levels of linguistic representations (e.g. syntax vs. lexicon) might rely differentially on implicit vs. explicit learning.
Despite potential advantages in metalinguistic awareness and applying conscious learning strategies compared to children, adult L2 learners in informal environments experience one considerable disadvantage, which children do not. Specifically, children, especially young children, are usually addressed in some form of child‐ or infant‐directed speech (originally termed motherese by Newport et al. (1977)). Child‐directed speech is characterized by exaggerated intonation contours (Vihman 1996), affectively linked intonation (Fernald 1992) that seems to be associated with speech acts across languages (Bryant and Barrett 2007), cleaner, more distinct vowels (Andruski et al. 1999), and cognitively and perceptually accessible in terms of the topics usually addressed (Zukow 1990), with more elaboration about meaning (Chouinard and Clark 2003; Gelman et al. 1998). In short, child‐directed speech serves to “carve up” the input into more and more discrete subsections, over which SL can more efficiently operate to extract generalizations.
In fact, narrowing or limiting of input might represent a significant difference between child‐language learning and adult L2 learning in both formal and informal environments. As adults, we assume our mature cognitive capacities allow us to take in and process more input, and that this increased capacity might be an advantage in L2 learning. Recent investigations into child‐language and L2 language learning, however, seem to suggest this assumption is wrong. It appears to be the case that less is more when it comes to learning from language input. Yang defines the Tolerance Principle (2016, 60ff.), which determines the ratio of items that are candidates for the application of a given rule over the number of exceptions to that rule. One surprising aspect of the Tolerance Principle, as Yang demonstrates mathematically, is that rules are easier to learn if they are defined over smaller sets of items. Children, thanks to a generally more limited range of input, i.e. child‐directed speech, which is more informatively presented, likely have a significantly more useful body of input to run their SL operations on. It is possible that adult L2 learners in informal learning contexts consider too much information when trying to extract rules and regularities from the flood of adult‐world input. As suggested by Yang's Tolerance Principle: Less might be more.
This said, perhaps adult cognitive capacities not only can handle a larger range of input than children can, but they also benefit from it more. Lev‐Ari (2018) reports intriguing results from a study applying social network analysis to the development of vowel perception in adult L2 speakers and children L1 speakers. In short, what is found is that adults benefit from broader, more heterogeneous language input from many speakers. In contrast, according to Lev‐Ari's computations simulations, young child L1 learners do not appear to benefit from more diverse input, and may in fact be hindered by it (although they do seem to learn better with input from more than one person). If these results are replicated in real people (as opposed to computational simulations), it may be that “less is more” for children, but “more is more” for adults in informal language learning environments, at least when it comes to phonological learning. It is interesting to think about how amount and diversity of input might have different effects at different levels of representation, as well, and how “optimal” input seems to differ for children and adults, depending on level of representation. As far as we know, this is an area that has not been addressed much in the literature.
So should adult L2 learners in informal environments eschew their cognitive prowess to try to return to some sort of “child‐like” state in order to more efficiently learn the ambient language? Although the benefits of doing so are not clear, this might not be a bad idea; however, it is likely not possible. Furthermore, along with Lev‐Ari's (2018) work, there is additional evidence that the cognitive capacities that adults can bring to bear on L2 learning are quite beneficial. For example, whereas children learn their native language(s) largely implicitly, adults are much better at explicit learning than children. Although the products of explicit learning by adults are more likely to fade over time than the products of children's implicit learning, adults can bolster learning and retention by leveraging their metalinguistic awareness of the patterns and material being learned (Smalle et al. 2017).
Adults are also better able than children to integrate multiple sources of information during language learning, which can help make up for less efficient SL processes. Take, for example, recent research by Lavi‐Rotbain and Arnon (2017), who followed up on work by Thiessen (2010), which showed that adults performed better in a word segmentation task in an artificial language when new lexical items were paired with consistent visual cues (objects). In Thiessen's study, infants did not benefit from the concurrent presentation of visual cues. Lavi‐Rotbain and Arnon similarly found that adult word‐segmentation performance benefited from visual cues, as did older children's performance on the same task; however, the performance of younger children (who already knew that words are used to label objects, unlike Thiessen's infants) was facilitated less by the concurrent visual cues. Taken together, these findings are paradoxical for our focus here. Although infants and young children seem to be worse than adults in integrating multimodal cues in the service of speech segmentation (a fundamental early step of language learning), they are nevertheless ultimately better than adults at speech segmentation in learning actual languages in informal (i.e. real‐life) environments. Whatever the source of this paradox, it is clear that adults can, and almost certainly should, take advantage of any environmental cue that helps them learn the L2. These cues might be systematic, such as iconic gestures paired with new L2 words, or they might be wholly idiosyncratic, such as arbitrary gestures paired with new L2 words, both of which have recently been shown to be effective in aiding word learning (but again, in relatively small batches; see Huang et al. 2019).
Finally, there is one difference between adult and child language learning that adults would be wise to eliminate, namely, resistance to production. Anxiety is detrimental to L2 learning (e.g. Macintyre and Gardner 1991), and anxiety about making mistakes or sounding foolish is associated with lower L2 proficiency (Dewaele et al. 2008). This sort of anxiety is likely not a frequent concern in L1 acquisition (although we know of no research into the topic). In L2 learning, though, it can severely reduce learners' L2 production. Adults might be prone to reason that they want to avoid L2 production until they learn enough via comprehension to attempt error‐free L2 speech. This reasoning seems to be deeply flawed, however. Recent work by Hopman and MacDonald (2018) suggests that increased production in a foreign language during learning improves subsequent comprehension. In other words, there appears to be a sort of feedback loop such that comprehension provides the raw materials for production, and then that production – even if errorful – facilitates later comprehension, not unlike the idea behind the long‐lag structural priming used by Shin and Christianson (2012).
Throughout history, there have always been people who possess exceptional facility for languages. These people are generally referred to as polyglots. Emil Krebs (1867–1930) was a well‐documented polyglot who, by the age of 45, was said to have spoken some 32 languages fluently and whose brain was preserved at the University of Düsseldorf (see Hyltenstam 2016). As reported by Kenneth Hyltenstam, Krebs's brain exhibited some atypicalities compared to control brains, most notably with respect to Brodmann areas 44 and 45 in both hemispheres. Precisely how these atypicalities related to his linguistic facility is not known, however.
Some polyglots' abilities appear to be associated with acute brain injury, rather than congenital atypicality. A famous “polyglot savant” is Daniel Tammet (1979–), who suffered a catastrophic epileptic seizure at the age of four. Along with prodigious mathematical abilities that developed subsequent to his first seizure, Tammet has mastered 11 languages, including Icelandic, which he reported to have learned in a week. Hyltenstam (2016) reports that when Tammet was tested on live TV in Iceland, he shocked the interviewers by carrying on a fluent conversation with them.
Exceptional cases like these notwithstanding, some polyglots have more typical etiologies, and are simply motivated for various reasons to learn multiple languages. Paradowski and Wysokińska (2014) conducted in‐depth interviews with six polyglots, each of whom spoke between 5 and 12 languages. Although a number of motivational theories were discussed, none of the six polyglots interviewed seemed to be driven by strictly integrative (“fitting in”) or instrumental (“utilitarian”) motives for learning. Moreover, all of them spoke languages that had been learned both in classrooms and in informal environments, with no appreciable differences in their ultimate attainment. Perhaps most striking was that none of them considered themselves to have a “gift” for language; rather, they all said it took considerable work and dedication, irrespective of the environment or motivation.
Any discussion of informal environments and language learning must include virtual environments, where today's generation of aspiring polyglots can find a rich amalgam of formal and informal instruction, community, and conversation partners. Anecdotally, “internet polyglots” are beginning to rival the accomplishments of more historically famous cases like Krebs. Take for instance Timothy Doner, who in 2012 was touted as “the world's youngest polyglot” (Huffington Post). At the time, it was reported that Doner could speak 23 languages, which he had taught himself over summer vacations and weekends. Doner's linguistic efforts include a YouTube channel in which he posts lessons in his various languages, and tips for connecting with conversations partners across the globe. As detailed by Michael Erard (2012) in his book Babel no more: The search for the world's most extraordinary language learners, the interconnected world today offers unlimited exposure to countless languages via the internet. People like Doner, who have recently been termed hyperglots, need not live in a cosmopolitan city to be surrounded by speakers of the languages they are interested in. Even if they live in a small monolingual English‐speaking town in, say, central Iowa, they can simply jump on the internet and find native speakers with whom to converse and from whom to learn. There is precious little research at present that analyzes how, or even if, informal virtual environments differ from informal face‐to‐face environments with respect to foreign language learning. Even critical evaluations of formal for‐profit language‐learning programs are still few and far between. Virtual informal environments, then, pose substantial opportunity not only for language learners, but also for researchers.
To summarize, adults who are learning an L2 in informal environments have substantial cognitive powers at their disposal, which can be advantageously deployed in the language‐learning task. These powers include efficient multimodal processing, explicit learning strategies, mature working memory capacity, metalinguistic awareness that continues to develop as more languages are encountered, and expert control over at least one language already. Despite all of these strengths, however, adult L2 learners generally fail to learn their L2 as fast or as well as children learn their L1. We have suggested here that the difference between child L1 learners and adult L2 learners is the former's greater facility at extracting statistical regularities from the input and internalizing these regularities (SL). We have furthermore suggested that, although SL is the driver of implicit learning, explicit attention to errors is likely crucial for adult L2 learning, especially in informal environments. Error‐driven learning may be an interface between explicit learning and implicit learning, and adults appear to be particularly adept at monitoring for errors. This facility, however, might be a hindrance to learning if error‐monitoring becomes so pervasive that the adult L2 learner hesitates to produce or engage in the L2 (in person or virtually) for fear of making mistakes.
Given the torrential nature of adult language input in informal environments, adults might be over‐eager to try to “sort out” all of the input at once. If learning equations that describe child‐language acquisition can be applied to adult L2 learning – and it is in fact accurate that “less is more” when it comes to extracting productive generalizations from the input – then the same might hold for adult L2 learners. Rather than diving into the L2 pool at the deep end, it would seem to be wise for the adult L2 learner to wade gently into the shallow end, soaking in the statistical regularities floating around in the input. As the learner wades more deeply, she or he can apply adult cognitive resources at each step, seeking out opportunities to practice (in person or remotely over the internet) with as many different interlocutors as possible, using rules as they are being learned, all the while reflecting metalinguistically on what those rules are and how they might be extended.
35.170.81.33