Introduction
Becoming bilingual involves learning the complex structures and uses of a novel language. Explaining how adults acquire new languages is also complex and involves understanding the role of multiple factors, such as the type of linguistic structure, the nature of the input, and learners’ linguistic, cognitive, and social profiles. One approach to piecing together this puzzle has been to examine specific factors before addressing questions about their complex interactions. For example, many researchers have focused on the acquisition of second language (L2) grammar by controlling variables, such as prior knowledge of the L2 (e.g., Stafford, Bowden & Sanz, Reference Stafford, Bowden and Sanz2012), in order to gain more precise insight into the role of independent variables, such as proficiency (e.g., White, Genesee & Steinhauer, Reference White, Genesee and Steinhauer2012). For both behavioral and neural studies though, testing hypotheses with natural languages often entails inherent confounds: for example, between proficiency and age of acquisition (e.g., Weber-Fox & Neville, Reference Weber-Fox and Neville1996). One manner of exerting tighter experimental control over variables and confounds is by using artificial linguistic systems (ALSs), which are language-like systems created by researchers for the purpose of addressing particular research questions (e.g., DeKeyser, Reference DeKeyser1997; Esper, Reference Esper1925). The current review provides a brief synthesis of ALS research targeting the neural basis of L2 grammar acquisition by considering the validity of different types of ALSs and the specific contributions made by ALS research to questions about age of exposure to the L2, type of exposure, and online L2 learning mechanisms, with suggestions for future research.
The validity of ALS research for L2
ALSs come in many forms (Ettlinger, Morgan-Short, Faretta-Stutenberg & Wong, Reference Ettlinger, Morgan-Short, Faretta-Stutenberg and Wong2016; Grey & Tagarelli, Reference Grey, Tagarelli, Phakiti, De Costa, Plonsky and Starfield2018), including: (a) artificial grammars, comprised of rule-governed sequences of elements that do not carry meaning; (b) artificial languages, comprised of meaningful nonce words whose use is based on linguistic properties of natural languages; (c) semi-artificial languages, based on a particular natural language into which new linguistic properties are incorporated; and (d) miniature languages, based on a particular natural language but pared down to a finite and small number of words and linguistic structures. Researchers have argued that there are many benefits to using ALSs (Ettlinger et al., Reference Ettlinger, Morgan-Short, Faretta-Stutenberg and Wong2016; Folia, Uddén, De Vries, Forkstam & Petersson, Reference Folia, Uddén, De Vries, Forkstam and Petersson2010; Grey & Tagarelli, Reference Grey, Tagarelli, Phakiti, De Costa, Plonsky and Starfield2018). First, ALSs allow researchers to control or even eliminate confounding variables in research designs: for example, different amounts of prior exposure to an L2. Second, they allow researchers to manipulate key variables, such as similarity of the ALS to participants’ first language (L1). Third, the ALS can be designed so that participants can reach high levels of proficiency in relatively brief amounts of time, which makes them amendable to longitudinal studies. Because of their characteristics and benefits, it has been argued that ALSs may “constitute ‘test tube’ models of natural language” (Morgan-Short, Steinhauer, Sanz & Ullman, Reference Morgan-Short, Steinhauer, Sanz and Ullman2012a, p. 934).
Although there are several benefits to ALSs, one must also consider an important potential limitation – their ecological validity for L2 learning. This may be a particular concern for non-meaningful ALSs, because they do not involve form-meaning relationships found in natural language (Sanz & Morgan-Short, Reference Sanz, Morgan-Short and Sanz2005; VanPatten, Reference VanPatten1994). Several neural studies, however, provide indirect evidence regarding the ecological validity of both meaningful and non-meaningful ALSs in that they elicit neural processes and correlates similar to those found in natural language studies (Friederici, Steinhauer & Pfeifer, Reference Friederici, Steinhauer and Pfeifer2002; Morgan-Short et al., Reference Morgan-Short, Deng, Brill-Schuetz, Faretta-Stutenberg, Wong and Wong2015; Petersson, Reference Petersson2004; Silva, Folia, Hagoort & Petersson, Reference Silva, Folia, Hagoort and Petersson2017). This evidence is further strengthened by neural studies that have found shared neural mechanisms within subjects for ALS and natural language (Christiansen, Conway & Onnis, Reference Christiansen, Conway and Onnis2012; Petersson, Folia & Hagoort, Reference Petersson, Folia and Hagoort2012; Tabullo, Sevilla, Segura, Zanutto & Wainselboim, Reference Tabullo, Sevilla, Segura, Zanutto and Wainselboim2013). Interestingly, a recent meta-analysis comparing neural activation for non-meaningful ALSs and meaningful linguistic systems found several converging areas of activation (Tagarelli, Shattuck, Turkeltaub & Ullman, Reference Tagarelli, Shattuck, Turkeltaub and Ullman2019). Thus, extant evidence suggests that ALS and natural languages at least have some shared neural mechanisms, lending credence to the ecological validity of ALSs.
Another important piece of evidence relevant to the ecological validity of ALSs for L2 learning is whether performance on ALS tasks is related to performance on L2 tasks. For non-meaningful ALS, evidence for a relationship between ALS and L2 learning has not yet been established (for null results in regard to the relationship between artificial grammar and miniature language learning, see Robinson, Reference Robinson2010; Robinson, Reference Robinson2005). For a meaningful ALS, however, Ettlinger et al. (Reference Ettlinger, Morgan-Short, Faretta-Stutenberg and Wong2016) showed that artificial language learning was moderately correlated to a range of L2 Spanish measures. Thus, both behavioral and neural research suggests that meaningful ALS paradigms may be valid models of L2 acquisition, whereas currently only neural research supports the validity of non-meaningful ALSs. Given the larger goal of accounting for successful L2 learning as informed by what learners can do with the language and not only what their brains can do, the current review will focus on research with the strongest ecological validation: that is, research using meaningful ALS paradigms. An important direction for future research, however, will be to further determine the ecological validity of meaningful and non-meaningful ALS paradigms and to consider methodological issues particular to ALSs, such as the use of control groups (Hamrick & Sachs, Reference Hamrick and Sachs2018).
L2 grammar acquisition as informed by meaningful ALS research
Meaningful neural ALS research has focused on specific questions regarding the acquisition of L2 grammar primarily with two brain-based methods. First, in order to examine the neural processes of L2, researchers have employed electrophysiological methods that provide a scalp-recorded measure of the brain's electrical activity related to a cognitive event (Luck, Reference Luck2014). Particular event-related potentials (ERPs) have been associated with L1 processing (Swaab, Ledoux, Camblin & Boudewyn, Reference Swaab, Ledoux, Camblin, Boudewyn, Luck and Kappenman2012), including (a) an N400 effect, which is a negativity commonly associated with lexical-semantic processing that occurs over centro-parietal regions of the scalp around 400 ms after the onset of a word; (b) a P600 effect, which is a positivity often associated with more controlled grammatical processing that occurs over centro-parietal regions around 600 ms after the onset of a word; and (c) a left-anterior negativity (LAN), which has been found around 200–400 ms after the presentation of a word and has been associated with more automatic grammatical processing, although its elicitation is variable across studies and may not always reflect automatic grammatical processing (Tanner & van Hell, Reference Tanner and van Hell2014; Tanner, Reference Tanner2019). Second, in order to examine the neural correlates of L2, researchers have used functional magnetic resonance imaging (fMRI), which reflects the hemodynamic response evoked by a neural event and thus provides a measure of neural activation (Huettel, Song & McCarthy, Reference Huettel, Song and McCarthy2004). For L1 grammar, neural activation is typically found in the inferior frontal gyrus (IFG), especially in Broca's area, as well as in the superior temporal gyrus (STG), the basal ganglia, and other cortical areas (Ullman, Reference Ullman, Connor-Linton and Fasold2006). The typical neural responses from L1 ERP and fMRI research often serve as a basis for comparison for L2 research. Although the current review also adopts this approach, as it is used in the research being reviewed, it is critical that future research reconsider the L1/native-speaker comparison (Bialystok & Kroll, Reference Bialystok and Kroll2017; Ortega, Reference Ortega2018) as L2 learners do not develop into native-speakers of the L2 but rather develop as bilinguals.
Age of exposure
Comparing L1 and L2 neural responses has served as a basis for addressing a central question in L2 regarding whether there is a critical, or sensitive, period for language acquisition (Lenneberg, Reference Lenneberg1976). This hypothesis was supported by early L2 behavioral (Johnson & Newport, Reference Johnson and Newport1989) and ERP (Weber-Fox & Neville, Reference Weber-Fox and Neville1996) research with natural languages that examined learners exposed to L2 at different ages. Even when late-learners with high proficiency in a natural language were examined, they demonstrated P600 effects but did not show the more automatic LAN response (Hahne, Reference Hahne2001).
In the first ERP study with an ALS, Friederici et al. (Reference Friederici, Steinhauer and Pfeifer2002) trained learners on the artificial language, Brocanto, to the highest levels of proficiency – 95% accuracy, which is possible to achieve in a reasonable amount of time with an ALS paradigm. After training, learners evidenced both P600 and LAN effects when processing Brocanto phrase structure. Subsequent ALS research with a miniature language based on Japanese, designed “to create an experimental design with a higher degree of ecological validity as compared to artificial language studies” (p. 1231), found that learners showed P600s for phrase structure and case when trained to 75% accuracy (Mueller, Hahne, Fujii & Friederici, Reference Mueller, Hahne, Fujii and Friederici2005). Earlier negativities for phrase structure, and for a subset of case violations after additional training to 95% accuracy (Mueller, Hirotani & Friederici, Reference Mueller, Hirotani and Friederici2007), showed a distribution across the scalp that differed from native-speakers. However, when tested on familiar words used in training, more native-like negativities followed by P600s emerged (Mueller, Reference Mueller2009).
More recently, fMRI ALS training studies have begun to address whether late-learned L2 grammar can rely on L1 neural correlates and mechanisms. For example, Nevat, Ullman, Eviatar, and Bitan (Reference Nevat, Ullman, Eviatar and Bitan2017) found activation in IFG for L2 morphology, and Finn, Hudson Kam, Ettlinger, Vytlacil, and D'Esposito (Reference Finn, Hudson Kam, Ettlinger, Vytlacil and D'Esposito2013) found that the recruitment of IFG predicted L2 learning success. Results from Morgan-Short, Deng, Brill-Schuetz, Faretta-Stutenberg, Wong, and Wong (Reference Morgan-Short, Deng, Brill-Schuetz, Faretta-Stutenberg, Wong and Wong2015) provided additional evidence for the involvement of IFG in L2 syntax, although the results also suggested that this may differ by individual. Finally, Weber, Christiansen, Petersson, Indefrey, and Hagoort (Reference Weber, Christiansen, Petersson, Indefrey and Hagoort2016) found that neural mechanisms operated on known ALS L2 grammatical structures as for known L1 structures. Thus, each fMRI ALS L2 study found results indicating involvement of L1 neural correlates and mechanisms in L2.
Taken together, these studies suggest that L2 learners have the ability to rely on native-like neural processes and correlates, although they may not always fully reflect those of native speakers. Interestingly, the pattern of neural results seems to parallel behavioral research with high proficiency L2 speakers of a natural language who were assessed on various tasks and who showed native-like performance on some but not all measures (Abrahamsson & Hyltenstam, Reference Abrahamsson and Hyltenstam2009). Overall, some L2 learners seem capable of native-likeness in their performance and in their recruitment of neural mechanisms, but not all learners reach such levels, and late-learners may not be able to evidence native-likeness on all assessments. Whether this pattern of results supports the critical period hypothesis is a matter of debate, but the neural research attests to at least some level of language-related neural plasticity at later ages.
Moving forward, neural ALSs could be leveraged to train L2 learners of different ages to examine potential age effects in the development of native-like processes and correlates, following behavioral research designs that examine L2 proficiency across different aged learners (Birdsong & Molis, Reference Birdsong and Molis2001; DeKeyser, Reference DeKeyser2000; Granena & Long, Reference Granena and Long2013). In addition, both behavioral and neural research should use bilingual populations as a comparison group (Bialystok & Kroll, Reference Bialystok and Kroll2017; Ortega, Reference Ortega2018). Finally, researchers should consider the role of type of exposure or training in achieving fully native-like neural patterns (Mueller, Reference Mueller2006).
Type of exposure/training
A second question addressed by meaningful ALS research is whether different types of exposure affect L2 grammar acquisition differentially. Ample behavioral research has examined this question and an important meta-analysis (Norris & Ortega, Reference Norris and Ortega2000) suggested that explicit learning contexts, which provided metalinguistic explanation or direction to attend to grammar, led to greater L2 gains than implicit contexts, which did not provide such explanation or direction, although implicit contexts also led to gains. However, the amount of exposure and the type of practice in primary research was not always controlled in these explicit and implicit comparisons, and it was not known whether these different types of exposure elicited different neural processes.
Adapting the Brocanto paradigm, Morgan-Short and colleagues (Morgan-Short et al., Reference Morgan-Short, Steinhauer, Sanz and Ullman2012a; Morgan-Short, Sanz, Steinhauer & Ullman, Reference Morgan-Short, Sanz, Steinhauer and Ullman2010) developed the artificial language Brocanto2 to control amount of exposure and type of practice and to establish grammatical structures (phrase structure and grammatical gender agreement) that differed from learners’ L1. Participants were provided with either explicit or implicit training on the language and were tested twice – once when they reached low proficiency during practice (45% accuracy) and once when they completed all training and practice (~90% accuracy). Results indicated that, for phrase structure, explicitly and implicitly trained learners did not differ behaviorally but developed different ERP signatures (Morgan-Short et al., Reference Morgan-Short, Steinhauer, Sanz and Ullman2012a): The implicitly trained group evidenced N400 effects at low proficiency and a LAN, P600, and late anterior negativity at the end-of-practice. The explicitly trained group only evidenced an anterior positivity and a P600 at the end-of-practice. At retention testing three to six months later, again no behavioral differences were found, but neural processing differences remained (Morgan-Short, Finger, Grey & Ullman, Reference Morgan-Short, Finger, Grey and Ullman2012b). For grammatical gender agreement, explicitly and implicitly trained learners also did not differ behaviorally (with the exception of a larger gain for noun-adjective gender agreement for explicitly trained learners) and showed both similarities and differences in their ERPs (Morgan-Short et al., Reference Morgan-Short, Sanz, Steinhauer and Ullman2010): At low proficiency, only the implicitly trained group evidenced N400s. At end-of-practice, both groups evidenced N400s for noun-adjective agreement and P600s for noun-determiner agreement. Overall, across both phrase structure and grammatical gender agreement, only one behavioral difference emerged between explicitly and implicitly trained groups, whereas several ERP differences were evidenced.
Batterink and Neville (Reference Batterink and Neville2013) also addressed the question of whether different types of exposure affect L2 neural processing. Learners were explicitly or implicitly trained on a miniature language based on French with a small set of grammatical structures (article-noun agreement, subject-verb agreement, and subject-verb-object word order). Results indicated behavioral advantages for the explicitly trained learners, although a subset of implicitly trained learners also reached high proficiency levels. For the ERP results, processing differences generally did not emerge between the groups when learners with similar levels of proficiency were compared: All learners at high proficiency evidenced P600s for all three linguistic structures, although explicitly trained learners also showed an early negativity for verb agreement.
Overall, the results from Batterink and Neville (Reference Batterink and Neville2013) are largely consistent with Norris and Ortega (Reference Norris and Ortega2000) and more recent behavioral meta-analyses (Goo, Granena, Yilmaz & Novella, Reference Goo, Granena, Yilmaz, Novella and Rebuschat2015; Spada & Tomita, Reference Spada and Tomita2010) in that explicit training led to better performance and more native-like neural processing, although implicit training also led to L2 learning and native-like processing. In contrast, results from the Brocanto2 studies (Morgan-Short et al., Reference Morgan-Short, Steinhauer, Sanz and Ullman2012a; Morgan-Short et al., Reference Morgan-Short, Sanz, Steinhauer and Ullman2010; Morgan-Short et al., Reference Morgan-Short, Finger, Grey and Ullman2012b) suggest that explicit and implicit training may not always lead to behavioral differences and that implicit training may lead to a fuller range of native-like ERP effects. The fact that the pattern of results differed between these ALS studies provides a motivation, first, for replication of each study in order to further establish the general validity of the results. Second, future research should examine the factors that differed between the studies, such as (a) the number of lexical items, (b) the pre-practice training, and (c) the amount, type, and spacing of practice, especially when it is theoretically motivated to look at these factors. Fortunately, ALS paradigms are quite conducive to these factors being examined independently and in interaction with other factors.
L2 online learning mechanisms
A third area in which neural ALS research is beginning to make important contributions pertains to our understanding of the online mechanisms that underlie the acquisition of L2 grammar. Note that the studies reviewed above examined neural processing and correlates on post-exposure or post-training assessments and not during learning itself. To the author's knowledge, only two meaningful ALS neural studies have examined L2 grammar online. Batterink and Neville (Reference Batterink and Neville2014) examined electrophysiological data that were recorded during training in the miniature French study (Batterink & Neville, Reference Batterink and Neville2013). Processing during training was analyzed in regard to ERP effects tied to selective attention (N100) and to successful identification of words (N400). Results implicated a role for selective attention during L2 learning as the N100 successfully predicted performance, whereas the N400 did not. In a different study, Batterink, Oudiette, Reber, and Paller (Reference Batterink, Oudiette, Reber and Paller2014) examined the processing of a semi-artificial language, adopted from Leung & Williams (Reference Leung and Williams2012; Reference Leung and Williams2014), with a grammatical rule that was present in the input but experimentally hidden. EEG data were recorded during training and during a nap that occurred half way through training. For learners who reported becoming aware of the rule during training, ERP analyses revealed a post-nap P600, which the researchers attributed to explicit processing. Interestingly, the duration of slow-wave sleep was associated with larger P600s. For learners who reported that they did not become aware of the rule, a post-nap negativity was found, which was attributed to implicit processing.
These studies suggest that the approach of using ALSs to examine the neural mechanisms involved during L2 learning may be particularly informative. Indeed, L2 research has moved in the direction of examining L2 learning mechanisms online using eye-tracking methods (e.g. Godfroid & Uggen, Reference Godfroid and Uggen2013; Godfroid, Reference Godfroid2019; Indrarathne, Ratajczak & Kormos, Reference Indrarathne, Ratajczak and Kormos2018; Issa & Morgan-Short, Reference Issa and Morgan-Short2019). Such research often focuses on the role of attention and awareness in L2 grammar learning. Thus, connections between eye-tracking and ERP methodologies related to these factors might provide unique insights to the mechanisms underlying L2 learning. Note that it would be quite difficult for cross-sectional studies or even longitudinal research with full natural languages to provide direct evidence about the mechanisms that support L2 learning because these paradigms invariably test L2 processing at particular points in time. In contrast, longitudinal ALS studies seem quite amenable to testing predictions about the neurocognitive mechanisms that play a role as learning occurs.
Conclusion
Research with meaningful ALSs has examined central questions about L2 grammar acquisition, and results suggest that (a) adult L2 learners can evidence native-like neural processes and correlates, (b) different types of exposure may affect L2 neural processes but do not necessarily do so, and (c) neural mechanisms related to attention may underlie online learning. Overall, using ALSs to test hypotheses in controlled models and then exploring these effects in interaction with other factors and in natural language seems to be an effective and productive way of developing knowledge about L2 acquisition and processing. With further corroboration of the ecological validity of ALSs, the implications of research with these paradigms for natural L2 may be further strengthened.
