Event related potentials at initial exposure in third language acquisition: Implications from an artificial mini-grammar study

The present article examines the proposal that typology is a major factor guiding transfer selectivity in L3/L n acquisition. We tested first exposure in L3/L n using two artificial languages (ALs) lexically based in English and Spanish, focusing on gender agreement between determiners and nouns, and between nouns and adjectives. 50 L1 Spanish-L2 English speakers took part in the experiment. After receiving implicit training in one of the ALs (Mini-Spanish, N = 26; Mini- English, N = 24), gender violations elicited a fronto-lateral negativity in Mini-English in the earliest time window (200 – 500 ms), although this was not followed by any other differences in subsequent periods. This effect was highly localized, surfacing only in electrodes of the right-anterior region. In contrast, gender violations in Mini-Spanish elicited a broadly distributed positivity in the 300 – 600 ms time window. While we do not find typical indices of grammatical processing such as the P600 component, we believe that the between-groups differential appearance of the positivity for gender violations in the 300 – 600 ms time window reflects differential allocation of attentional resources as a function of the ALs ’ lexical similarity to English or Spanish. We take these differences in attention to be precursors of the processes involved in transfer source selection in L3/L n .

structure to be the motivator of transfer selection, but it differs by claiming that transfer is holistic (full transfer in the sense of Schwartz & Sprouse, 1996), obtaining from the earliest moment the parser receives enough linguistic information to determine overall structural proximity. As a result, this full transfer model advocates for an initial interlanguage grammar that is fully specified from its outset, that is, once the full copy of an underlying system has been established as the initial interlanguage of the L3.
With the above context in mind, Rothman, Alemán Bañón, and González Alonso (2015) advocated a methodological expansion in formal linguistic approaches to L3/Ln morphosyntactic studies, sitting at the cross-roads of neurolinguistics and artificial mini-grammar learning. At the time, use of electroencephalography (EEG) in formal L2 morphosyntax had been providing promising results suggesting that representational transfer effects could be captured with event-related potentials (ERPs). Work showing evidence of very early qualitatively-similar-to-natives processing for properties available for transfer, juxtaposed against the lack of such effects when transfer is not possible, was amassing, e.g. number versus gender processing in English learners of Spanish (Alemán Bañón, Fiorentino, & Gabriele, 2018;Bond, Gabriele, Fiorentino, & Alemán Bañón, 2011;Gabriele, Fiorentino, & Alemán Bañón, 2013). At the same time, evidence of qualitatively similar processing for gender violations after reaching advanced proficiency was showing that properties not part of the L1 could be acquired (Alemán Bañón et al., 2018). If acquired, L2 representations would in principle be available for transfer in multilingual acquisition/processing. Showing that ERPs can capture transfer effects and that learners can/do attain native-like signatures for grammatical processing meant that the precursors for applying this method were in place. Rothman et al. (2015) argued that an ERP extension into L3 morphosyntactic transfer studies was ideal because such data register involuntary reactions at the brain level with excellent temporal resolution. They are thus not subject to metalinguistic variables in the way offline behavioral methods could be. Moreover, since ERPs provide a window of sorts into the inner working of the mind as parsing and processing occurs in real-time, EEG data could shed a unique light onto the process of transfer selection itself, especially if done very early on in the process. As such, ERPs would be as welcome an addition to the formal linguistic L3 morphosyntactic literature as they had been for the L2 literature. At the same time, Rothman and colleagues argued that employing an ERP method under the conditions of artificial mini-grammar learning would serve an additional purpose, beyond the obvious one related to capturing evidence at the very earliest stages. Artificial mini-grammar learning allows for maximal control of crucial variables, not least the quantity and the qualitative shape of language exposure. Working with semiartificial or fully artificial grammar learning is not new to non-native language research (e.g., Morgan-Short, Steinhauer, Sanz, Ullman, 2012;Rebuschat & Williams, 2012;Williams & Kuribara, 2008). In the majority of such cases, however, this paradigm is used to examine implicit versus explicit learning. Rothman et al.'s reason for using it was somewhat distinct. Using two artificial mini-grammars, as described below, offered two methodological advantages that would have been nearly impossible to achieve outside such a context. First, it allowed for control of the exact quantity and context of exposure to the L3 across all participants, ensuring everyone was at a very beginning proficiency level and received maximally comparable input. Second, it allowed for manipulation of the qualitative shape of the input the participants received, taking advantage of lexical similarities to either the L1 (Spanish) or the L2 (English) of the participant groups to make two artificial mini-grammars. As we will explain below, this was a crucial step in testing between default models and those that advocate structural similarity as the main motivator for transfer selection, in particular the TPM (the only such model existing at the time of Rothman et al.'s (2015) paper).
Using mini grammars in L3/Ln research is rare (e.g. Grey, Williams, & Rebuschat, 2014;Sanz, Park, & Lado, 2015) and, to date, has investigated similar questions pertaining to implicit versus explicit learning as when applied to L2. Combining ERP and a (semi) artificial grammar paradigm in non-native language research is a bit more common, although the vast majority of this regards adult L2 acquisition (see Morgan-Short, 2020 for review). To our knowledge, the few studies (e.g., Grey, Sanz, Morgan-Short, & Ullman, 2018) that have combined ERP with artificial grammar learning in L3 contexts have done so for distinct questions to those we pursue here. However, findings from Grey et al.'s (2018) study are highly relevant. They examined behavioral and neural correlates of learning an artificial language in early Mandarin-English bilinguals, compared to English monolinguals as a control. After grammar instruction, participants were tested by means of a grammaticality judgement task at low-and high proficiency while EEG was recorded. Most interesting to the present context are the low proficiency results, where bilinguals and monolinguals did not differ on behavioral measures, but showed distinct ERP patterns. Only the bilinguals showed a P600, a common ERP correlate of combinatorial processing, for sentences that were ungrammatical in the artificial grammar. Grey et al. (2018) did not align their results to the L3 literature as presented above, as they wanted to test the claim that bilinguals are "better" learners of subsequent languages (Cenoz, 2003). And so, the present study is not only the fulfillment of the discussion and methodology presented by Rothman et al. (2015), but remains the first ERP paper to weigh in on the formal linguistic L3 literature.

Research questions
The fundamental research questions guiding this study are: RQ1. What linguistic and extra-linguistic variables constrain the selection of a transfer source in non-native language acquisition, when previous language experience instantiates multiple (and potentially conflicting) options?
RQ2. (How) Can neurolinguistic measures of implicit language processing, such as those provided by EEG/ERPs, help us to adjudicate between the competing predictions of several models developed on the basis of behavioral data?
In particular, we test for the first time if the sensitivity of ERPs to (potentially) different neurocognitive underpinnings of converging behavior can permit us to distinguish between performance rooted in transferred linguistic representations vs. performance based on a variety of learning/task strategies (rhyming, pattern matching, etc.).

Predictions
With these questions in mind, we first follow the predictions of Rothman et al. (2015) with respect to the three main models existing at the time (the L2 Status Factor, the CEM and the TPM) for potential ERP effects for the processing of gender violations in two artificial mini-grammars, lexically based on Spanish (Mini-Spanish) and English (Mini-English), respectively, by different groups of Spanish-English sequential bilinguals. The key feature of these semiartificial systems, as originally conceived by Rothman et al., was that both would display gender agreement between nouns and adjectives. Number agreement was used as a control, because both Spanish and English realize it within the NP-although only Spanish displays noun-adjective agreement. In contrast, grammatical gender applies only to Spanish. Every Spanish noun is inherently assigned masculine or feminine gender. Adjectives and determiners display a matching systematic morphophonological agreement pattern. English lacks this feature altogether.
In this connection, it is worth noting that the relationship between the P600 and agreement violations in native speakers is not a one-to-one correspondence. The P600 can be elicited by non-linguistic events (e.g., disruptions in musical harmony, Patel, Gibson, Ratner, Besson, & Holcomb, 1998;visual degradation, van de Meerendonk, Chwilla, & Kolk, 2013), suggesting that it might reflect sensitivity to the structural configuration of rule-based knowledge beyond (linguistic) syntax. Also, it sometimes appears in response to linguistic anomalies that are typically subsumed under semantics (e.g., Gunter, Friederici, & Schriefers, 2000;Kuperberg, 2007;Kuperberg, Caplan, Sitnikova, Eddy, & Holcomb, 2006). Taken together, these findings led to a reinterpretation of the P600 as reflecting late integration of syntactic, semantic and thematic information (see Friederici, 2017, pp. 62-65, for an overview and discussion). Furthermore, agreement violations have been reported, in a subset of studies, to elicit other types of electrophysiological responses, such as a biphasic pattern where the P600 is preceded by a negativity in the 200-500 ms time window, often with a left-anterior distribution, known as the LAN (e.g., Friederici, Hahne, & Mecklinger, 1996;Molinaro, Barber, Caffarra, & Carreiras, 2015). 2 Non-native speakers display more variability in their electrophysiological responses to agreement violations, which vary as a function of factors such as L2 proficiency and L1-L2 combination (see, e.g., Alemán Bañón et al., 2018). N400 responses without a subsequent positivity, which are rare in sentential contexts in the native speaker literature, can be found in low-proficiency L2 speakers for agreement violations of both shared (L1-L2) and novel (L2 only) features (e.g., Carrasco et al., 2017;Osterhout, McLaughlin, Pitkanen, Frenck-Mestre & Molinaro, 2006;Tanner, McLaughlin, Herschensohn, & Osterhout, 2013), which tend to elicit more native-like, positive-dominant responses at higher proficiency (e.g., Alemán Bañón, Fiorentino, & Gabriele, 2014Foucart & Frenck-Mestre, 2012; see also Rossi, Kroll, & Dussias, 2014, where the novel feature, gender, only elicited native-like responses at very high proficiency). The rate of development to native-like processing of agreement dependencies, however, is arguably faster for non-novel features (i.e., those already present in the L1; e.g., Gabriele et al., 2013), for which P600-like responses have been reported even at low levels of proficiency (e.g., Alemán Bañón et al., 2018;Alemán Bañón, Hoffman, Covey, Rossomondo & Fiorentino, under review;Tokowicz & MacWhinney, 2005;cf. Osterhout, McLaughlin, Pitkänen, Frenck-Mestre, & Molinaro, 2006;Morgan-Short, Sanz, Steinhauer & Ullman, 2010, who report negative-dominant responses at this stage). Crucially for our purposes, the fact that prior experience with certain linguistic features can result in early native-like electrophysiological indices suggests that linguistic transfer can be captured and meaningfully teased apart from other potential sources of target-like behavior in grammatical agreement comprehension (for example, phonological associations between word pairs, which would yield qualitatively different ERP components, such as an N400, as shown by studies manipulating rhyming; e.g., Coch, Hart, & Mitra (2008) from the earliest stages of non-native language development.
The relevant predictions in Rothman et al. (2015) are schematized in Table 1. Let us walk through the reasoning behind the predictions appearing in the table. The predictions of the L2 Status Factor and the CEM are the same for both participant groups (L1 Spanish learners of L2 English) regardless of which artificial mini-grammar they are exposed to, Mini-Spanish or Mini-English because they relate to aspects that are fixed across either the participant groups or the sum total of previous language experience for both. In other words, the L2 is the same for both groups, and the property of interest (grammatical gender) can also be found in only one of the previously acquired languages (Spanish; incidentally the L1) in both cases. The L2 Status Factor would predict that L2 English is the default source of transfer across the board, which, for gender, means that the 2 The LAN does not appear in all relevant studies, leading some to argue that when it obtains it is more likely a by-product of averaging across within-group individual differences related to negative-dominant vs. positive-dominant responses to agreement violations (e.g., Guajardo & Wicha, 2014;Tanner, 2019;Tanner & Van Hell, 2014; see also ;Grey, Tanner, & Van Hell, 2017;Osterhout, 1997;Osterhout, McLaughlin, Kim, Greenwald, & Inoue, 2004; but see; Caffarra, Mendoza, & Davidson, 2019). Nevertheless, the P600 is indeed consistently found for agreement violations when group-level grand averages are considered (see Molinaro et al., 2011, for a comprehensive review).
L3 Mini-Spanish and Mini-English grammars should not transfer a full specification for this property. For this reason, the ERP prediction derived in Rothman et al. (2015) was that neither group should show sensitivity, as reflected in electrophysiological measures, to gender violations. The CEM makes the opposite prediction, given that transferring grammatical gender from Spanish would provide a target-like representation in the L3. Since both groups have the property in their previous linguistic experience, then, they both should show sensitivity to gender violations, reflected in an ERP pattern that is qualitatively (though almost certainly not quantitatively) similar to those elicited by grammatical gender violations in L1 (dominant) Spanish speakers (e.g., Alemán Bañón et al., 2012): a P600, potentially preceded by an N400/LAN (Molinaro, Barber, & Carreiras, 2011).
The TPM makes distinct predictions. Each group is expected to perform differently as a consequence of the artificial mini-grammar to which they are exposed, despite having the same linguistic backgrounds. The TPM claims transfer is conditioned on overall structural similarity, that is, that either the L1 or the L2 grammar is copied in full as the initial L3 interlanguage grammar based on how much structural crossover is estimated between them and the incoming L3 input as early as possible during initial exposure. This is captured by the TPM's processing hierarchy. The parser is argued to use a set of linguistic cues in a particular order to assess the relative similarity between the L3 and each of the previously acquired grammars. This cue hierarchy takes into consideration the differential salience of various elements in the linguistic input, and the relative precedence that some of these elements must logically take over others (e.g., it is simply not possible to parse functional morphology, or even basic word order, without at least some knowledge of the words involved in a given sentence). The first of these levels is the lexicon, followed by phonology/phonotactics, functional morphology and, finally, syntax (see Fig. 1).
In our experiment, the artificial mini-grammars of Mini-Spanish and Mini-English are, in fact, identical in morphosyntactic structure (see the Methods section for details): They have the exact same word order and the exact same morphosyntactic properties of gender and number agreement with the exact same morphophonological exponents. However, their lexical bases are radically dissimilar, as they are entirely based on Spanish and English, respectively. From the perspective of the TPM's processing hierarchy, then, being exposed to Mini-Spanish vs. Mini-English represents a completely different task. The very high percentage of lexical overlap between the L3 and one of the previous languages only (the L1 in the case of Mini-Spanish, the L2 for Mini-English) should bias the selection of a transfer source in favor of the lexically closest language (Spanish or English), irrespective of the presence or absence of grammatical gender in that language-and, crucially, despite it featuring in both artificial mini-grammars. In short, L1 Spanish L2 English speakers exposed to Mini-Spanish should, after exposure training, transfer Spanish as the initial L3 interlanguage, and English should be transferred for those speakers exposed to Mini-English. In terms of ERP responses, Rothman et al. (2015) suggested this should translate into a sensitivity to gender violations in the form of P600-type components (potentially as a biphasic response, preceded by an N400) only for the Mini-Spanish group, with L3 Mini-English learners showing no sensitivity to these violations.

Artificial mini-grammars
Two artificial mini-grammars (henceforth ALs [artificial languages], for convention) were created, based on the English (Mini-English) and Spanish (Mini-Spanish) lexicons, respectively (e.g., Grey et al., 2014;Rebuschat & Williams, 2012;Williams & Kuribara, 2008). Each contained 12 nouns (six masculine, six feminine), 12 adjectives, one article (four allomorphs: masculine singular, feminine singular, masculine plural, feminine plural), a copula (two inflected forms: singular and plural), one coordinative conjunction, one adverb and two locatives. Adjectives carried number (singular/plural) and gender (masculine/feminine) morphology in both languages. These morphemes, as well as the article, were common to both ALs, and thus phonotactically plausible in both English and Spanish. Gender and number information was marked on both the article and the adjective, but not on the noun. Gender assignment Table 1 Original ERP predictions for the processing of gender violations in Mini-Spanish and Mini-English by L1 Spanish-L2 English speakers, derived from the L2SF, CEM and TPM (adapted from Rothman et al., 2015).
was inherent to each noun, congruent with the gender of the equivalent nouns in Spanish. These were selected so that half of the nouns in each gender display transparent assignment in Spanish (feminine nouns ending in -a and masculine nouns ending in -o), and half have nontransparent assignment (ending in -e or a consonant). Mini-Spanish nouns were stripped of -a and -o endings, and any original diacritics or exclusive characters (accents, ñ) were removed or substituted by approximate characters that were nonexclusive (i.e., shared by the Spanish and English alphabets; e.g., ñ > n). Table 2 contains a full list of stimuli in both ALs.
To facilitate implicit training through adjectival contrasts (see below), adjectives consisted of six pairs of opposing qualities (e.g., old/new, expensive/cheap). The adverb and conjunction were introduced in order to construct spillover regions after the critical violations in the ERP experiment. The locatives were used in filler sentences, to reduce the predictability of an adjective continuation after the copula.
Number morphology consisted of an -r affix to mark plural adjectives and articles, akin to Spanish and English -s. Gender morphology was built on a consonantal contrast, (− )j-(feminine) vs. (− )z-(masculine), carefully selected to avoid consonants with strong associations to functional morphology in either source language (e.g., n-as a recurrent segment in negation-related morphology). The use of a consonantal instead of a vowel contrast was necessary to achieve this same objective (all vowels have associations to one or several functional morphemes in Spanish) as well as to reduce the opportunity for potential misleading rhyming strategies with the -e endings of some of the nouns. Allomorphs of the article and the inflectional affix on the adjective had the following structure: In the adjective, the distinctive consonant (j or z) was invariably followed by the letter -e. In the adjective, the epenthetic vowel -e-preceded the distinctive consonant, which was invariably followed by the letter -u. The different vocalic endings on the article and the adjective were designed to avoid the development of rhyming strategies in the learning of the ALs' gender morphology.
To isolate the lexical similarity factor in accordance with the implicational hierarchy of the TPM, all typological differences between Spanish and English that may have been potentially inherited by Mini-Spanish and Mini-English were methodologically neutralized. First, our study employed visual (as opposed to auditory) training and presentation of stimuli, in order to remove all possible phonological bias. Second, gender and number morphology were common to both ALs and compliant with each of the source languages' phonotactics. Finally, the experiment examined morphosyntactic violations on the marking of predicative adjectives following a copulative verb, a structure in which English and Spanish word order coincides (as opposed to the pre-/postnominal contrast displayed by most attributive adjectives in these two languages). These manipulations ensured that the only difference between Mini-Spanish and Mini-English lay on their lexical base, while both languages instantiated gender agreement.

AL training
A pre-training phase familiarized participants with the nouns, by presenting a randomized list of picture-phrase combinations containing two instances of each noun in simple article-noun phrases (e.g., je window). This phase thus introduced both the lexical material and the nouns' lexical gender.
The critical items in the ensuing training session exploited adjectival contrasts to provide implicit exposure to number and gender morphology through meaningful sentences in the AL without metalinguistic explanations. These sentences were presented as Table 2 Full AL design and vocabulary items, and example sentence (feminine singular). mochil, taz, ventan, pared, llave, calle. cuchil, gor, roper, camion, reloj, lapiz bag, cup, window, wall, key, street knife, hat, closet, truck, watch, pencil Adjectives Adjectives amarill-, roj-, pequen-, grand-, nov-, vej-, suci-, limpi-, barat-, car-, cort-, larg-yellow-, red-, small-, big-, new-, old-, dirty-, clean-, cheap-, expensiv "The bag is cheap." Je bag is cheapegu. "The bag is cheap." descriptions of pictures showing the object and property denoted by the noun and the adjectives, respectively (see Fig. 2 for an example). The training session lasted approximately 45 min. In order to maintain an implicit focus on the morphology, none of the critical items in the training session contained either number or gender contrasts, relying instead exclusively on the semantic differences encoded in the adjectives. That is, an image depicting a masculine or feminine plural object was always contrasted to an image depicting the same object in the plural-and likewise with singular items. To encourage mapping of the prenominal element to the intended functional categories (i.e., singular and plural articles instead of numeral determiners), half of the images depicting plural nouns contained two objects, and the other half contained three. The training contained a total of 144 critical items, resulting from crossing the 12 nouns and 12 adjectives. Two different lists were created to counterbalance the presence of a given noun-adjective combination in singular or plural form. The training stimuli were completed with 144 fillers-for a total of 288 items-using the locative prepositions (e.g., The bag is below the hat).
At the end of the training, the participant's command of the AL was tested through a multiple-choice sentence-picture matching task. Upon presentation of a single picture (e.g., three cheap bags), the participant was asked to select the sentence that best described the image among five alternatives: correct (e.g., Mini-Spanish Jer mochil son baratejur), a sentence containing a gender violation (*Jer mochil son baratezur), a sentence containing a number violation (*Jer mochil son barateju), a sentence containing a double violation (*Jer mochil son baratezu) or a sentence containing only a semantic violation (that is, the opposite adjective with the right morphology, e.g., #Jer mochil son carejur). The test consisted of a total of 36 items, in which knowledge of all possible gender and number morphology was assessed.

Participants
A group of 60 Spanish native speakers (52 female), L2 speakers of English, were recruited as participants from three different areas of South-East England (Southampton, London, Reading) and tested at the University of Reading. Completion of a short questionnaire ensured that they met the study's language background requirements (Spanish as a mother tongue, English as a late-acquired L2). The age range was 18-51 (mean = 28.5; SD = 8.61). Their average proficiency in English, as measured by the Oxford Quick Placement Test (Oxford University Press et al., 2001) was 40.1 for the Mini English group (SD = 11.08; 95% CI = 4.63) and 35.7 (SD = 8.32; 95% CI = 3.74) for the Mini Spanish group, thus placing them between lower intermediate and advanced at the individual level. Importantly, the groups did not differ in their English proficiency (t(43) = 1.60; p = 0.12), and all participants were proficient enough to have acquired English number agreement, the only L2 grammatical property involved in the design. Average age of acquisition of English was 11.08 (range: 0-26; SD = 7.92). Participants were randomly assigned to one of the two AL groups, which meant that they completed the training and ERP experiment (if eligible, see below) exclusively in that language (i.e., they are never exposed to the other AL). Therefore, all participants had Spanish as their L1, English as their L2, and either Mini-English or Mini-Spanish as their L3. All participants were neurotypical adults with normal or corrected-to-normal vision, and were right-handed according to the Edinburgh Inventory for the Assessment of Handedness (Oldfield, 1971). Participants were compensated for their time.

Procedure
Upon arrival, participants were informed of the general procedure of the study, provided written consent, and completed a short questionnaire that gathered demographic and language background data. Next, they were assigned to an AL group and seated comfortably in a quiet room, in front of a 32-inch monitor. After being exposed to the pre-training and training phases described above, the participants were tested through the sentence-picture matching task. If they scored above 80% in this task (equivalent to 29 correct responses out of 36 total), they moved on to the ERP experiment. Otherwise, they were re-exposed to the training. Failure to reach the 80% accuracy threshold on a second take of the test resulted in the participant's removal from the study. This resulted in the exclusion of less than 10% of the original participants (N = 4).
After the ERP experiment was completed, participants performed a gender assignment task (GAT) in Spanish that contained the 12 nouns which served as the lexical base in Mini-Spanish (and were translated as the 12 critical nouns in Mini-English) together with 108 fillers, for a total of 120 items. This was done to ensure that gender assignment in our participants' lexical representations of these nouns matched the intended values. In the task, each noun was presented in isolation on the center of the screen, and participants were asked to respond, by pressing one of two buttons on the mouse, matching the noun to the appropriate determiners. A short set of six trials familiarized participants with the procedure before the critical and filler items were presented in a single block of trials. Accuracy and response times (RTs) were collected for each response, although there was no timeout in the GAT and participants were encouraged to favor accuracy over speed. As expected (given that these were all native speakers of Peninsular Spanish), results showed full alignment across the board with the expected gender assignment for the critical nouns.

ERP experiment
In the ERP experiment, participants read sentences on a computer screen while electrophysiological activity was recorded at the scalp. Sentences were presented word by word, employing the rapid serial visual presentation (RSVP) method. Each word appeared on screen for 450 ms, with an inter-stimulus interval of 250 ms. After the last word of the sentence, a prompt with a happy and a sad face on each side of the screen indicated that a grammaticality judgement was required. Using their right hand index and middle fingers, participants responded by pressing the left or right button on a mouse. While response times were recorded as an additional behavioral measure, there was no time limit on the grammaticality judgement.
The ERP experiment contained the same 144 critical sentences as the training, resulting from crossing the 12 nouns and 12 adjectives. Three versions of each sentence were created: grammatical, gender violation and number violation. No double violations (i. e., gender and number) were ever present in the stimuli. Table 3 contains a sample of the conditions in the Mini-English and Mini-Spanish experiments, up to the critical word. Violations were always on the adjective, which represented the critical word in our analyses. All critical sentences contained a final region consisting of the conjunction (and/y) plus another determiner phrase (DP) of the same gender and number as the sentence's initial DP and the adverb too/tambien (e.g., je bag is cheapegu and je cup too; Mini-English, grammatical) (Alemán Bañón et al., 2012Bañón et al., , 2014Bañón et al., , 2018. This extra material was introduced to prevent any potentially confounding wrap-up effects from interfering with ERP responses elicited by the critical word. Each participant saw a total of 240 sentences: 96 critical sentences (48 gender violation, 48 grammatical) plus 144 fillers (48 number violations and 96 grammatical fillers). The grammatical fillers followed the same structure as those in the training session (DP + copula + locative + DP), with a final region akin to that of critical sentences (e.g., Zer truck are below zer closet and zer watch too). Different experimental lists counterbalanced the conditions in which the different noun-adjective combinations appeared, as well as the number of times in which each initial DP was used in the singular or plural form. The 240 trials were divided into six blocks of 40. Participants were encouraged to rest their eyes and take as long as they needed before continuing with the experiment.

EEG recording and analysis
The data from six participants was lost due to technical problems. Together with those excluded for not having reached the accuracy threshold in the sentence-picture matching task, this left a total of 50 participants: 24 in the Mini English group, 26 in the Mini Spanish group.
The EEG signal was continuously recorded from the scalp by means of 64 active electrodes (ActiCap, Brain Products, Inc.) fitted in an elastic cap in a 10-20 system. AFz served as the ground electrode. EEG recording was referenced online to electrode FCz, and rereferenced offline to the average mastoids (TP9/10). The fronto-parietal electrodes FP1 and FP2, located above the eyebrows, were used to monitor eye blinks. Impedance was kept below 10 kΩs for all electrodes. The recordings were amplified by a BrainAmp MR Plus amplifier (Brain Products, Inc.) with a bandpass filter of 0.01-200 Hz, and digitized continuously at a sampling rate of 1 kHz.
Preprocessing of the EEG data was performed on Brain Vision Analyzer 2.0 (Brain Products, Inc.). We used all trials, regardless of accuracy in the GJT (e.g., Mickan & Lemhöfer, 2020;Tanner et al., 2013;VanRullen, 2011). First, data were filtered with a band-pass filter of 0.1-30 Hz. The continuous EEG was then segmented into 1500 ms epochs with reference to the critical word (adjective). Epochs contained a 300 ms pre-stimulus baseline and ended 1200 ms after stimulus onset. Trials were manually inspected for artifacts (drifts, excessive muscle artifact, blinks, blocking, etc.). Rejecting trials with artifacts resulted in the exclusion of 43.3% trials in the Mini-English data, and 24.7% trials in the Mini-Spanish data. 3 The average number of trials kept did not differ by condition within each AL (Mini English, Number: 26, Gender: 27, Grammatical: 29; Mini-Spanish, all conditions: 36; all ps > .1). The remaining epochs were baseline-corrected relative to the 300 ms pre-stimulus baseline and averaged by condition for each participant.
The mean amplitudes for electrodes in these regions were entered as dependent variables into factorial repeated-measures ANOVAs for each time window of interest. Separate analyses were carried out for lateral and midline electrodes (e.g. Alemán Bañón & Rothman, 2016;Barber & Carreiras, 2005; Gillon Dowens, Vergara, Barber, & Carreiras, 2010). ANOVAs for the lateral electrodes included Condition (grammatical, gender violation), Hemisphere (left, right) and Caudality (anterior, medial, posterior) as predictors, whereas the models for the midline included only Condition and Caudality. In those cases in which the assumption of sphericity was violated (as determined by Mauchly's test) for a given main effect or interaction, we report adjusted degrees of freedom and p-values following from corrections based on Huynh-Feldt epsilon estimates. While this correction is less conservative than Greenhouse-Geisser estimates, it is particularly appropriate for data of this kind, where the nature of the study (ab initio learners, artificial mini-grammars) is likely to yield relatively large between-and within-subject variability. Follow-up analyses on any significant interactions were conducted by means of one-way repeated-measures ANOVAs on the appropriate subsets of the data.
All analyses of both the behavioural and the ERP data were conducted in the statistical software R (R Core Team, 2019), using the packages lme4 (Bates, Maechler, Bolker & Walker, 2015; for linear mixed models), multcomp (Hothorn, Bretz, & Westfall, 2008; for post hoc comparisons of effects in the linear mixed models), and ez (Lawrence, 2016; for repeated-measures ANOVAs).

Behavioural data
Accuracy and response time (RT) data were collected for the participant's responses to each grammaticality judgement. Table 4 provides a summary of descriptive statistics. After removing the data from filler sentences, the raw accuracy scores and RTs were submitted to statistical analysis.
We fit a generalized linear mixed model (binomial family) on the accuracy data, with condition (grammatical, gender) and AL as fixed factors (main effects and interaction), and the maximal random structure justified by design (Barr, Levy, Scheepers, & Tily, 2013). This included random intercepts for subjects and items (in our case, the critical adjective) as well as random slopes for condition within subjects and within items. There were no main effects or interactions of AL (all ps > .4). Overall, responses to gender violations were only numerically less accurate (β = 0.71; z = 1.93; p = .053).
For the RT data, we fit generalized linear mixed models with an inverse Gaussian family. We began by fitting a maximal model that included condition and AL as fixed factors (main effects and interaction), as well as random intercepts for subjects and items and random slopes for condition within both. We proceeded to decorrelate random slopes and intercepts and to remove random slopes accounting for the least amount of variance until a model achieved convergence. This final model included random intercepts for subjects and items. There were no main effects or interactions of either AL or condition (all ps > .4). In sum, while the behavioral data show certain numerical trends, with overall higher accuracy in grammatical sentences across both ALs, there were no significant differences in accuracy or RTs as a function of condition in either AL group.

ERP data
We report here only main effects or interactions with the Condition factor (that is, we leave out main effects or interactions which involve exclusively Caudality and Hemisphere). The full output of the omnibus ANOVA for each time window in each AL can be found in Appendix A (Tables A1 to A12).

Mini-English
Topographical maps and grand averaged ERP waveforms for the Mini-English group can be seen in Figs. 3 and 4.

Table 4
Mean accuracy (in percentage) and RTs (in milliseconds), with standard deviations and 95% confidence intervals, for each condition in the two ALs.

Condition
Accuracy (

Mini-Spanish
Topographical maps and grand averaged ERP waveforms for the Mini-Spanish group can be seen in Figs. 5 and 6.

Summary of ERP results
Gender violations elicited a fronto-lateral negativity in Mini-English in the earliest time window (200-500 ms), although this effect was not followed by any other differences in subsequent periods. This effect was highly localized, surfacing only in electrodes of the right-anterior region. In contrast, gender violations in Mini-Spanish elicited a broadly distributed positivity that starts trending in the 200-500 ms time window, emerges significantly between 300 and 600 ms, and dissipates in the later period.

Discussion and conclusion
The results described above present an interesting scenario; one that does not conform to any expectations in light of current theories, yet captures a difference between how the AL groups perform. It is thus important to first acknowledge that these results are unexpected and discuss how they deviate from the predictions above, and then proceed to engage with the data themselves in order to interpret what it is they tell us about the earliest stages of L3 processing. Thus, related to our original research questions, we note that the data are rather limited with respect to adjudication between the L3/Ln models we originally sought to test. Nevertheless, they offer rich insights as it pertains to the perennial question of what is happening at the very beginning of L3 exposure, thus potentially addressing what factors feed into eventual transfer selection itself, the process as opposed to its outcome.

Revisiting the original predictions
Put simply, the ERP predictions derived above from the TPM, the CEM, and the L2SF were not substantiated by the present data. Recall that, as an extension of their different claims regarding the (under)specification of initial L3 interlanguage representations, we stated that (i) the CEM would predict both groups to show sensitivity to gender violations that is qualitatively similar to that of native speakers (a P600 or biphasic N400-P600 response), because both would transfer the linguistic representation for gender in Spanish; (ii) the L2SF would predict that both groups should lack this sensitivity and therefore show no effects in response to gender violations, because they would always transfer from their L2 (English in both cases); and (iii) the TPM would predict only those participants exposed to Mini-Spanish to show sensitivity (in the form of a P600 or a biphasic N400-P600 response) to gender violations, since the model predicts that English should be transferred as the initial interlanguage grammar for L3 Mini-English-thus lacking this property-and, conversely, a copy of the Spanish grammar should become the initial L3 interlanguage grammar for Mini-Spanish-thus instantiating gender agreement (see Table 1 for ERP predictions).
A quick glance at the results shows that the data did not conform to the scenarios described in (i)-(iii). The early, broadly distributed positivity in the Mini-Spanish group does not resemble the P600 that should be associated with the transfer predictions of the CEM or the TPM, and neither does it validate the no-effects prediction of the L2SF. 4 At best, the effects of the Mini-Spanish group could be consistent with Steinhauer, White and Drury's (2009) claim that non-native learners will initially show a broad P600 (across the scalp), which then becomes focalized. We will return to this below. Similarly, the early right anterior negativity displayed by the Mini-English group was not anticipated by any of the models. Importantly, we know of no current theory, including those not addressed in detail here (e.g., the LPM), that would have predicted the current pattern of results. It is important to keep in mind, however, that all predictions were predicated on the assumption that testing would have happened after transfer would have taken place. Thus, the onus was on examining the outcome and using it for evidential adjudication. However, if indeed the current methodology did not allow sufficient time (from initial exposure to testing) to get at the outcome stage, as we will argue occurred below, then the data are not only still valid but potentially crucial, albeit at a different level: examining the process of how selection is conditioned in the first place.

ERP indices of pre-transfer processing?
Although other mini grammar studies have claimed to show evidence of transfer after minimal exposure, this has been the case in L2 studies where there is no competition between more than one system. In L3 acquisition, by comparison, it proved in our case to be problematic to assume that transfer selection among choices would take place in a similar time course. No transfer model offers a rubric of timing for selection-even the most articulated current models of L3/Ln acquisition offer only indirect estimations of the timing of transfer. The TPM's cue hierarchy, for example, provides an implicit rubric of timing that is ultimately dependent on overall L1/L2 to L3 similarity (typology, in a broad sense): moving through the levels of the hierarchy requires increasing exposure/time, which means that transfer will inevitably be delayed in language triads where the L3 bears no immediately apparent resemblance to the L1 or the L2 in, at least, vocabulary and/or segmental and suprasegmental phonology.
It follows from the above reasoning that the inverse relation should also hold: (full) transfer should be accelerated in language triads where the L3 has a strong similarity at the lexical level with either the L1 or the L2. This was in fact the rationale for our design, where the ALs hide identical underlying grammars under a heavily Spanish-or English-biased lexicon. Given this similarity, which is only disrupted by the introduction of novel agreement morphology, we should have been able to tap into an early L3 interlanguage grammar after relatively brief AL exposure. However, there was no guarantee that our training phase, which consisted of about 45 min 4 Even under an L2SF account where the Declarative Procedural (DP) model is argued to underlie the basis of L2 default selection (see Bardel & Falk, 2012), predictions are not supported. Under such an account, which assumes that L2/Ln grammatical representation is subserved by declarative memory, one might predict N400s for gender errors (as the DP model takes modulations of this component to reflect the involvement of the declarative memory system; Ullman, 2016), but this is not seen in either group. under implicit learning conditions, would not fall short of the actual amount of exposure required to trigger any kind of representational transfer from a previous language. Given this possibility, we must consider an additional scenario with a corresponding research question, extending the original scope of Rothman et al. (2015): What should we expect to observe in terms of brain evidence, according to the different models of transfer in L3/Ln acquisition, if transfer has not (yet, or at all) taken place? Do they differ at all? We think so.
Framing the above in the context of the current study is not an easy task, as none of the transfer models has explicitly dealt with these questions in the published record. Nonetheless, there are certain predictions that must ultimately follow from the different conceptual frameworks of these theories. In particular, default models of L3/Ln transfer, for which order of acquisition (i.e., L1 vs. L2) is the most deterministic factor in transfer source selectivity, have grounded this position on the basis of processing arguments related to a relative prominence of the L1 or L2 as a first-pass filter to the incoming L3 input. The L2 Status Factor, in particular (see Bardel & Falk, 2012), assumes Paradis' (2009) Declarative/Procedural (DP) model, where the L1 grammar is argued to be subserved by procedural memory while all other linguistic material, including L1/Ln lexicon and L2/Ln grammars are subserved by declarative memory systems. In other words, the L2 Status Factor attributes the L2's purportedly dominant influence on L3 grammar building to the fact that any subsequent grammar to those acquired natively is dealt with by the same neurocognitive subcomponents, in all cases distinct from the L1. Since the L2 should be guiding all handling of the L3 input, it seems reasonable to assume that the L2SF would expect electrophysiological activity reflecting an English bias in both groups.
The TPM, on the other hand, does not envision a predetermined, input-independent privileged role for either previous language upon initial exposure to an L3/Ln. On the contrary, the mechanics of the cue hierarchy and the principled approach to a best-guess estimation of overall structural similarity between the L3 and previously acquired languages have been explicitly argued to capture a "sink or swim" reaction to the task of parsing the incoming L3 input (see Chapter 4 in Rothman et al., 2019, for a comprehensive discussion). In this sense, the specific linguistic properties of the input are crucial: any property of the input that allows the processor to adjudicate between potentially conflicting parsing strategies should be capitalized on. Lexical similarity is likely to be a powerful trigger-and thus stands at the top of the TPM processing hierarchy. In our current design, this should translate into the English-like vocabulary of Mini-English biasing the parser in favor of processing strategies that focus resources on those aspects of sentences that are critical to the processing of English, and the Spanish-like vocabulary of Mini-Spanish biasing the allocation of processing resources in ways that are necessary when parsing Spanish sentences. While there is little theoretical apparatus to base conjectures on for other non-default models, such as the CEM or the LPM, we assume here that a similar logic would apply to them in these very early stages.
We believe that the current results are more in line with the latter type of accounts. The first, and perhaps most obvious, reason is that default models based on order of acquisition, irrespective of whether they envision a privileged role of the L1 or the L2, seemingly preclude the possibility of different behavior across our two groups, which are identical populations exposed to different L3 ALs. The second is that these divergent responses to gender agreement violations in Mini-Spanish and Mini-English are expressed through differences in ERP patterns that suggest our participant groups' attentional resources were differentially allocated, and this affected their perception of the critical manipulation in our experiment. In particular, we submit that the early positivity in the Mini-Spanish group's response to gender violations is reminiscent of the P300 (or P3), a well-known ERP component that is thought to reflect the recruitment of attentional and memory resources (e.g., Polich, 2012). The P300 is a (family of) positive going wave(s), elicited by both auditory and visual stimuli that are motivationally significant (Sassenhagen, Schlesewsky, & Bornkessel-Schlesewsky, 2014). Its latency and amplitude vary greatly as a function of different manipulations in stimulus frequency and task-relevance (e.g., Luck, 2014;Polich, 2012;Woodman, 2010), and different subcomponents have different topographical maxima (frontal and parietal being the most common; Luck, 2014). While first observed within the oddball paradigm (Squires, Squires, & Hillyard, 1975), where a series of frequent stimuli is occasionally interrupted by an infrequent deviant stimulus (see, e.g., Donchin, 1979Donchin, , 1981, for early reviews), it was quickly established that the P300 could not be indexing an estimation of absolute frequency, as it was elicited by task-relevant stimuli even when the frequency distributions of different stimulus categories were balanced in a given experiment (e.g., Duncan-Johnson & Donchin, 1977). In our study, the ratio of deviant (violations) to frequent stimuli (grammatical sentences) can be considered as low as 1:4 or as high as 2:3, depending on what is counted as a distinct stimulus category-i.e., depending on whether number violations, fillers here, are included in the deviant stimulus count. In any case, and since participants were explicitly asked to evaluate the grammaticality of these sentences, violations can be considered task-relevant stimuli. It is worth noting that our effect is broadly distributed and thus does not seem to present either a frontal or a parietal maximum, but it is difficult to ascertain with the present data whether this could be related to an averaging effect of different overlapping subcomponents-whereby the nature of our study (artificial mini-grammars, first exposure) may have increased the weight of anterior effects, typically linked to novelty (e.g., Squires et al., 1975;Verleger, Jaskowski, & Wauschkuhn, 1994), counterbalancing the parietal bias present in most P300 studies employing a standard oddball paradigm.
As pointed out by Sassenhagen et al. (2014:30), the fact that the P300 is sensitive to deviations that are not infrequent suggests that the component reflects subjective rather than inherent salience. In fact, some studies have found that unattended stimuli do not elicit a P300 (Nieuwenhuis, Aston-Jones, & Cohen, 2005;Spencer, Dien, & Donchin, 2001)-not even the frontally maximal P3a subcomponent associated with task-nonrelevant stimuli ("ignore" stimuli) in the seminal paper by Squires et al. (1975). Since our Mini-Spanish participants show this positivity and our Mini-English participants do not, it is reasonable to posit that only the former are attending to the relevant properties of the deviant stimuli-i.e., the gender violations-and that they do so under the influence of a larger focus on word-final gender morphology that a Spanish bias should induce.
While our focus here has been on gender, precisely because this is the grammatical feature not shared across Spanish and English, if our hypothesis is on the right track converging evidence from the number condition is relevant. While both English and Spanish have syntactic number, morphological number agreement on the adjective-herein we targeted both gender and number in the EEG recording for both mini-grammars on the adjective-only pertains to Spanish. And so, it is not clear what the effect of having number as a syntactic feature alone, as is the case in English, would translate to under our current hypothesis; however, the insight is that it would not necessarily matter because there is no word final number morphology on English adjectives. For the case of the Spanish minigrammar group we would expect a similar P3 effect as was shown for gender. Indeed, visual inspection of the data from the Mini-Spanish group shows a broadly distributed early positivity (300-500 ms), which is followed by a later more focused positivity over typical P600 sites, although occurring later than the typical P600 time window (starting at 700 ms). This contrasts with the mini-English group, which shows a late negativity (500-900 ms) through canonical N400 sites, though a positivity emerges around 300-500 ms in frontal sites-dissipating before 700 ms.
Before delving deeper into the nature of the positive component in the Mini-Spanish group, we should devote some attention to the negativity found in the Mini-English results. Indeed, besides the absence of the positivity that we interpret as a P300 in the other group, responses to gender violations in the Mini-English group seem to be characterized by a very localized right anterior negativity in the 200-500 ms time window. There are no precedents, to the best of our knowledge, of a comparable component in language-related studies. Music processing research has identified a replicable ERP known as the ERAN (early right anterior negativity), reliably elicited by structural violations such as disruptions of musical scale (e.g., Bharucha & Krumhansl, 1983;Koelsch, Schmidt, & Kansok, 2002;Leino, Brattico, Tervaniemi, & Vuust, 2007;see Koelsch, 2009, for review), normally in conjunction with later components. The ERAN was only reported in one study employing concurrent musical and linguistic stimuli, and there the presence of syntactic violations in conjunction with harmonic expectancy violations seemed to reduce the amplitude of the component (Steinbeis & Koelsch, 2007). In linguistic research, Hahne and colleagues have observed right anterior negativities in nonnative speakers (e.g., Hahne, 2001;Hahne & Friederici, 2001), but these obtain (i) in auditory experiments; (ii) in later time windows (e.g., following an N400), and (iii) in response to distinctly semantic violations. It might be tempting to think that, in comparison with the potentially attention-related positive component shown by the Mini-Spanish group, the early right anterior negativity of the Mini-English participants reflects an automatic detection of the gender violation that is otherwise unattended. However, even in studies that report an early negativity in response to unattended structural violations, this is either markedly left-dominant (Batterink & Neville, 2013) or broadly distributed (Hasting & Kotz, 2008), in which case it has a significantly earlier onset (100 ms vs. 200 ms) and shorter latency (200 ms vs. ~350 ms) than the early right anterior negativity present in our data. We are, in short, unable to find a functionally credible account of the Mini-English negativity (if it is indeed a real effect), although we believe that there are sufficient reasons to dismiss the possibility that it indexes any attentional or anomaly-detection processes. Crucially, however, it is of significant value to point out that whatever underlies the difference between the groups, their brainwave patterns are distinct despite being equal in all other relevant ways concerning language experience and other profile factors. Thus, this effect must have something to do with the discrepancy in the stimuli they were given, which we controlled to be at the lexical level in relation to one or the other language (Spanish or English) they have matched knowledge on. This means that the language to which one is exposed in L3/Ln acquisition matters for determining how one initially parses/processes it above and beyond order of acquisition of previous languages.
If the positive component in the Mini-Spanish group is indeed an instance of the P300, it seems meaningful to ask whether this could be a precursor of more identifiable language-related responses at later stages-i.e., after longer exposure to the L3 ALs. We have no verifiable way to know, but there are reasons to highlight this as a valuable (and viable) inquiry for future research. A healthy amount of studies have suggested that the P600 may in fact be a member of the P3 family (e.g., Bornkessel-Schlesewsky & Schlesewsky, 2015; Coulson, King, & Kutas, 1998;Sassenhagen & Bornkessel-Schlesewsky, 2015;Sassenhagen et al., 2014;cf.;Osterhout, McKinnon, Bersick, & Corey, 1996;Osterhout & Hagoort, 1999), given similarities such as the fact that the P600 is also absent or reduced for unattended relevant stimuli (e.g., Batterink & Neville, 2013;Hahne & Friederici, 2002;Hanulíkova, van Alphen, van Goch, & Weber, 2012) that it also peaks around centro-parietal sites, and that it seems to be response-time aligned (Sassenhagen & Bornkessel-Schelewsky, 2015;Sassenhagen et al., 2014). In this view, the later onset of the P600 relative to a standard P300 would be explained by the higher complexity of linguistic material as compared to the visual stimuli typically found to elicit a P300. Indeed, increasing the complexity of stimuli in a P300 experiment reliably delays the onset of this waveform (see Luck, 2014, pp. 98-101, for review).
What, then, could we expect of this component in terms of its developmental sequence, if it is indeed tied to the allocation of attentional resources that may over time result in the detection of syntactic anomalies? Studies examining the processing of non-novel agreement features in low-proficiency L2 learners have reported incipient P600-like effects, which already have the expected onset and scalp distribution but are much reduced in amplitude (Gabriele et al., 2013, under review). However, even the low-proficiency learners in Gabriele et al. (2013) had significantly more exposure to the target language than our participants. It is, therefore, possible that our data capture a much earlier stage where linguistic representations are underspecified, yet processing of the L3/Ln input is already guided by selective focus on those elements of the linguistic material that are most relevant to constructing reliable syntactic parses in one of the previous languages-e.g., affixes in Spanish, which typically contain agreement morphology necessary to process different kinds of long-distance dependencies. To the extent that this is not grammatical processing per se yet and, as such, evaluations of deviancy in the stimuli may not involve higher-order (linguistic) processing, an earlier onset of this positive component is not unexpected.
In such a short training exposure, it does make sense that we would tap processes of economizing for preparing an individual for the task of multilingual acquisition. And so, what we believe we are seeing is what the mind does in the most initial sense to get into the real process of subsequent linguistic acquisition, even before transfer takes place, by first finding the simplest ways to justify transfer selection among choices while still performing the task at hand. The simplest way to detect differences between grammatical and ungrammatical sentences would be to focus on the morphology and really only the morphology. In this sense, the participants in this task do not seem to be processing entire phrases per se, but rather some are detecting patterns that happen to be linguistic ones at a much smaller level. Crucially, however, only one of the two groups is doing this, indicating that something more general about the input one but not the other group receives determines whether smaller-unit pattern matching is activated to try to parse the incoming input stream. Given our manipulation of the mini-grammars, lexical similarity stands out as the only obvious candidate for the difference noted, which is compatible with models of transfer that focus on structural similarity as the main conditioning factor (such as the TPM and the LPM).
The above discussion highlights the importance of future research into the earliest stages of L3/Ln processing. Future work with longer training sessions, we believe, would determine the amount of exposure needed for one to have a better chance at capturing an initial true L3 grammar, thus allowing one to adjudicate between competing models of transfer source selection as was our original goal. Such studies have the potential to uncover the very dynamic nature of the processes that lead to early transfer in the first place, which are themselves connected to important questions regarding the implementation of cognitive economy in language. In fact, longitudinal studies making use of multiple testing and mid-term consolidation have the best chance to capture transitional stages that illuminate not only the mechanisms that establish initial specifications in the new (L3) grammar, but also potential missing links between the neural correlates of sensory and higher-order processing within a single developmental sequence. If it turns out that the P300 here reflects more focused attention at a stage before representational transfer has taken place-the selection phase-because the target language of exposure is lexically based on Spanish, then we would predict that in a design where there is a consolidation phase, re-exposure in an additional (or more) phase(s) and subsequent rerunning of the EEG/ERP we would see a P600 emerge. Such a design is not uncommon in the implicit vs. explicit non-native learning literature that employs similar designs for distinct questions (e.g., Morgan-short, Finger, Grey, & Ullman, 2012). We would hypothesize that under such a design the Rothman et al. (2015) predictions would be in the best position to be tested.