Second language (L2) acquisition of many aspects of syntactic structure is known to be difficult, especially when acquisition starts later in life. A major question being debated in the literature is to what extent and under what circumstances late L2 speakers can become native-like with respect to syntax processing (e.g., Clahsen and Felser, 2006; White, 2007). The evidence is mixed; in some cases this does seem to be possible, while in other cases, it is difficult or impossible. A number of factors have been suggested to play a role in this variation, but two which have received relatively little attention are the difficulty of the target grammatical system and the potential role of modality of testing (written vs. auditory presentation). The present study investigates whether event-related potential (ERP) measures of native-likeness used in this line of research might be partially dependent on stimulus modality, as this might explain some of the inconsistency in the literature.

A structure that has frequently been used to test native-like attainment in the L2, is grammatical gender, since it has been shown to pose a major challenge to L2 learners (e.g., Hawkins, 2001; White et al., 2001; Sabourin, 2003; Blom et al., 2008). Demonstrating gender processing that is comparable to that of natives therefore forms a strong test for L2 syntax acquisition. Grammatical gender is a classification system for nouns (e.g., masculine and feminine in French, or masculine, feminine and neuter in German) which allows speakers to establish syntactic cohesion between the elements in a phrase through agreement. Because the gender of a word is typically not predictable from its meaning, learning grammatical gender involves acquiring both the knowledge of a word's gender (gender assignment) and of how gender is expressed syntactically (gender agreement or concord). Therefore, L2 learners must tag each new lemma with its corresponding gender and learn which grammatical elements in the context have to agree with it. For example in Dutch, all nouns are assigned to either the common or the neuter gender class and gender concord occurs with determiners and pre-nominal adjectives (e.g., de_{[def, common]} tuin_[common], the garden, een_[indef] mooie_{[indef, common]} tuin_[common], a beautiful garden). During processing, a comprehender must retrieve the noun's gender fast enough to establish gender concord. The question is (a) whether L2 learners manage to do so, and (b) whether they achieve this using the same processing strategies as native speakers.

Gender processing in L2 has already been the topic of numerous investigations using behavioral measures, such as grammaticality judgments, sentence-picture matching, (elicited) production, and eye tracking (for overviews, see, e.g., Grüter et al., 2012; Hopp, 2013). More recently, researchers have begun to employ ERPs to investigate native-likeness of grammatical gender processing in the L2, because ERPs are known to be highly sensitive to the immediate, unconscious on-line detection, and processing of linguistic anomalies (e.g., Osterhout and Holcomb, 1992; Molinaro et al., 2011). Studies using off-line behavioral measures (e.g., White et al., 2001, 2004; Franceschina, 2005) cannot give access to this sort of evidence, which makes interpretation of their results more difficult. Some online techniques such as eye tracking (Dussias, 2010) measure real-time language processing, but do not provide us with the qualitative evidence of potential brain mechanisms that ERPs can. The rationale of such ERP studies is that the more similar the response between native speakers and learners, the more similar the underlying neural and cognitive processing mechanisms. In other words, a comparison of ERPs in native speakers and L2 learners can tell us how native-like the latter really are.

In first language processing, gender and other (morpho)syntactic violations are found to be associated with two primary kinds of components: the left anterior negativity (LAN) and the P600. The LAN has been widely associated with morpho-syntactic agreement processes (Münte et al., 1993; Friederici et al., 2000; Molinaro et al., 2011), but others claim that it is a more general index of working memory load (Kluender and Kutas, 1993; Coulson et al., 1998). The P600 has been reported for a range of syntactic and other linguistic violations (e.g., Osterhout and Holcomb, 1992; Hagoort et al., 1993; Münte et al., 1993; Burkhardt, 2007). Given the extremely heterogeneous conditions that elicit a P600, this component cannot be exclusively associated with agreement specifically, or even syntactic processing difficulties more generally, and is therefore often interpreted as a late stage of (re)analysis of information (Osterhout and Holcomb, 1992; Bornkessel-Schlesewsky and Schlesewsky, 2008). It may even reflect a more general process, such as the P300 (Gunter et al., 1997; Coulson et al., 1998; but see Osterhout and Hagoort, 1998; Frisch et al., 2003). There is however, a strong correlation between the appearance of the P600 effect and grammatical violations. In contrast, findings are more varied with respect to the presence of a LAN. In addition to the LAN and P600, some studies have found an N400, or a biphasic N400-P600 pattern (but no LAN) in response to syntactic violations (see an overview reported in Molinaro et al., 2011). This is surprising, since the N400 is a component normally associated with difficulty in semantic integration (see Kutas and Federmeier, 2011, for an overview). It has therefore been proposed that an N400 in response to syntactic agreement anomalies is likely to be a result of non-syntactic information that is needed to process the mismatch, for example information that requires lexical access (Molinaro et al., 2011). Because the LAN and N400 are variable in studies of native processing, particularly for gender agreement, we will consider the P600 to be the primary measure of native-likeness, although we will report findings in the time window associated with the LAN/N400 (300–500 ms after presentation) as well.

ERP results regarding grammatical gender processing in the L2 have provided mixed results. A number of studies find that, at least under some conditions, sufficiently proficient L2 learners are able to show native-like ERP responses to gender violations. A set of studies investigating L2 processing of French suggests that English, German, and Spanish learners of French can show native-like ERP responses in the form of a P600 effect (Frenck-Mestre et al., 2009; Foucart and Frenck-Mestre, 2011, 2012). The same goes for English and Chinese learners of Spanish (Tokowicz and MacWhinney, 2005; Gillon Dowens et al., 2010, 2011). German and Polish learners of Dutch can also show a P600 in response to gender violations (Sabourin and Stowe, 2008; Loerts, 2012). Despite these consistent results, however, it is clear that this does not generalize to success in all aspects of gender processing, as the English and German learners also failed to respond in a native-like manner to gender in some forms of agreement (Foucart and Frenck-Mestre, 2011, 2012). Stronger yet, Romance learners of Dutch did not show sensitivity to gender agreement anomalies in the form of a P600 effect even in straightforward determiner noun agreement structures (Sabourin and Stowe, 2008). It is unclear why this group failed to exhibit the majority pattern; we will discuss some factors which might have affected their success in somewhat more detail.

One of the factors which has been considered to be central for native-like learning of a late L2 is whether a grammatical element (e.g., gender) is present in the L1. Many studies have focused on this question, but have reached different conclusions. There is some evidence that having a gender system in the L1 might be an advantage when acquiring an L2 gender system (e.g., Bruhn de Garavito and White, 2000; Hawkins, 2001; Franceschina, 2005). This is in favor of models proposing that the L1 restricts L2 acquisition (Hawkins and Chan, 1997). However, there is also evidence of L2 learners without gender systems in their L1 being able to show full acquisition of grammatical gender (White et al., 2001, 2004), which is seen as evidence against such a restriction (Schwartz and Sprouse, 1994, 1996; see also White, 1989; White et al., 2004). The presence vs. absence of gender in the L1 seems at the least to be more complicated than these views suggest, however.

The French and Spanish studies mentioned earlier show that learners with no gender in their L1 (English and Chinese speakers) can show native-like ERP responses. Further, Sabourin and Stowe (2008) find differences between two L1s which both have gender: German on the one hand and Romance learners on the other. Sabourin and Stowe themselves attribute their results to the (lack of) similarity between the native and target language of these learners: Dutch gender is in general predictable from the gender of the cognate German word due to their common historical origin, while there is no one-to-one-correspondence between Romance and Dutch gender at the lexical level. Moreover, agreement between noun and adjective is more similar in German and Dutch than the Romance languages and Dutch. Sabourin and Stowe conclude that processing routines are transferred from L1 to L2, rather than transfer of the abstract knowledge that nouns have gender, and that these routines must be similar for successful transfer (see Foucart and Frenck-Mestre, 2011, for a similar argument).

However, an explanation which assumes that similar routines in L1 are necessary for native-like processing does not account for the results of other studies mentioned above showing that even with no gender system in the L1, learners are able to show native-like effects. A different approach to the effects of L1 transfer is formulated within the Competition model (see Bates and MacWhinney, 1987). According to the competition-based account, when L1 does not contain gender there is no interference. This predicts successful outcomes for languages with no gender (Tokowicz and MacWhinney, 2005). However, when existing processing routines are transferred, they will cause interference if they are dissimilar from those required for L2 (accounting for the failure of the Romance learners of Dutch).

The target language itself may also contribute to the failure of Sabourin and Stowe's (2008) Romance group to show native-like processing. Most of the successful studies have investigated Romance target languages. Unlike Romance or Slavic languages, which have transparent gender systems (i.e., a predictable gender category based on morphophonological patterns), Dutch is generally regarded as having an opaque gender system (Corbett, 1991; van Berkum, 1996). Although some morphological forms predict the gender of the word, these cues are only available for a relatively small proportion of the vocabulary in the language. This clearly presents a more difficult problem for the learner than gender in a more transparent language, which may certainly explain why the Romance group in the Sabourin and Stowe study failed to achieve a native-like level.

Neither L1 interference nor target language opaqueness, however, entirely accounts for the results found by Loerts (2012). Her study demonstrates that highly advanced Polish learners of Dutch can show somewhat weak, but native-like ERP responses, even though Polish agreement differs from Dutch. Loerts' results also show that an opaque system can be learned, although it may be more difficult to learn than a transparent system. Only her most proficient learners showed native-like processing (see Davidson and Indefrey, 2009, for another example of relatively low proficient learners failing to show native-like effects for gender processing in an opaque L2 system), while even fairly low proficient English learners of Spanish have been shown to respond with a clear P600 effect (Tokowicz and MacWhinney, 2005). An alternative explanation is thus that Sabourin and Stowe's (2008) Romance learners were simply not proficient enough to show online processing comparable to that of natives. Although the proficiency of the Romance group was not investigated in detail, a similar group of German learners did significantly better when tested on offline gender knowledge (Sabourin, 2003). The Romance participants in the ERP study also performed worse at the end of sentence grammaticality judgments collected during the ERP session. It has been shown that proficiency affects brain responses (e.g., Steinhauer et al., 2006; McLaughlin et al., 2010). A replication of the Sabourin and Stowe study with a group of learners as proficient as in the Loerts study can demonstrate whether this is the sole explanatory factor. This is one of the aims of the current study.

However, there is another factor that may have produced the difference between the two Dutch studies, which has thus far been overlooked: testing modality. Unlike virtually all the other studies summarized above, Loerts (2012) tested her Polish learners using auditory sentence presentation. She argues that the learners had acquired their L2 primarily in the auditory modality as emigrants who arrived with no formal training in their new language. Consequently, processing routines may be tuned to the auditory stimulus modality. Indeed, the experience of learning in immersion can be expected to differ substantially from a formal learning environment. Yet, the various populations that have been tested so far differ in this domain. The participants in the Romance studies summarized above included learners with extensive formal training in their L2. In many of the studies there was no immersion (Tokowicz and MacWhinney, 2005; Gillon Dowens et al., 2011) or only minimal immersion during the participants' recent residence in France (Foucart and Frenck-Mestre, 2011, 2012). Sabourin and Stowe (2008), unlike Loerts, tested a similar late immersion population using visual materials, with each word presented consecutively in the center of the screen. An alternative explanation for the lack of a native-like response in their study could thus be difficulties with the visual presentation. Below, we will speculate about why a visual ERP paradigm might, under some circumstances, be problematic.

In a typical language comprehension ERP paradigm, participants are presented with sentences displayed one word at a time at the center of a screen, at a rate of around two words per second, a technique called rapid serial visual presentation (RSVP). The advantages of this method are that the duration of stimulus presentation can be controlled (and manipulated) tightly, that eye movements, which lead to large artifacts in the EEG, are reduced to a minimum, and that making the stimulus material and time-locking the brain responses to the presentation of violations in the stimulus is relatively straightforward. Consequently, a large majority of ERP sentence comprehension studies use this method. In contrast, auditory sentence presentation is used much less frequently in ERP research. With spoken stimuli, it is more difficult to control the presentation duration of individual words. In addition, making recordings of spoken sentences is more time consuming and requires tight control of acoustic confounds (e.g., prosodic cues about upcoming information, Dimitrova et al., 2012), as well as timing issues (e.g., setting markers to millisecond precision for the events of interest).

We do not expect to find interesting differences between word-by-word reading and listening for language processing in natives (Müller et al., 1997; Hagoort and Brown, 2000; Balconi and Pozzoli, 2005). In the L1, learners develop fully automatized processing of both modalities; moreover, the auditory representation of language is automatically activated by written materials (Perfetti et al., 1992; Frost, 1998), so that the routines activated during auditory processing can be utilized as well as those specific to the written modality (Homae et al., 2002). Despite expecting comparable results for the two modalities in general, even for L1 comprehenders, consecutive word by word presentation in the middle of the screen presents a challenge under some circumstances. The optimum speed of presentation is an issue; Hopp (2010) shows that speeded RVSP presentation can make even native speakers break down in their grammaticality judgment ability, making their performance mirror that of L2 learners (see also Camblin et al., 2007, who show a case where speeded RSVP eliminates an effect which is clear in naturally produced connected speech). Conversely, studies directed at optimizing computerized text presentation on small screens have shown that too slow a presentation can also interfere with comprehension (Bernard et al., 2001). This may result from working memory and maintenance issues. Stowe (1991) showed that readers were more likely to garden-path or have difficulty in recovering from a garden path with center of the screen presentation, as opposed to presentation of words across the screen in their normal position, even when readers were allowed to pick their optimum pace.

L2 learners differ in a number of ways from native speakers, some of which can be expected to interact with modality. First, their cumulative reading experience in the L2 is likely to be substantially lower than that of native speakers. This means that their activation of the L2 via this modality can be expected to be less automatized than in native speakers (Koda, 1996). Second, interference from the writing system of the first language may lead to even less activation of the phonological form of the L2, in comparison with natives (Koda, 1999). These differences can potentially play a role for all L2 learners, but may be especially relevant for learners with less formal instruction in the language and in whom learning took place primarily via the auditory modality. The optimum speed of presentation is also likely to differ between various groups of learners and natives. This issue has received relatively little attention in the literature, but given that stimulus modality was one difference between the unsuccessful Romance group reported by Sabourin and Stowe (2008) and the relatively more successful group studied by Loerts (2012), this factor was included in the current experiment in order to determine whether it explains the different patterns seen in the two studies. A clear effect of modality would suggest that researchers need to pay more attention to this variable in their experimental designs, and might have implications for the differences between immersed and instructed learners as well.

Summarizing, the goal of the current study is to gain more insight into why some groups may show persistent problems in attaining native-like processing of grammatical gender. We investigate grammatical processing in immersed Romance L2 learners of Dutch, using the P600 as a measure of native-likeness, in order to answer the question whether late L2 learners can show native-like syntactic processing, even if the gender marking in the L1 differs from that in the L2, which may cause interference, and the L2 gender system is relatively opaque, making it harder to recognize the grammatical agreement regularities. Following Sabourin and Stowe (2008), in addition to gender violations, which have proven difficult to master, we present our participants with non-finite verb violations, a construction that is relatively easy to acquire, as a baseline for comparison. We compare the responses of high-proficient Romance learners with those of native speakers of Dutch. Additional measures of proficiency will be gathered from the first. A within-subject comparison of stimulus modalities allows us to determine whether the absence of a P600 effect for gender in the Sabourin and Stowe (2008) study was due to processing demands associated with the task modality.

In addition to standard group analyses of the ERP waveforms, we will closely inspect individual differences within each group. Adding these analyses has several benefits. First, lack of effects in grand mean ERP results does not necessarily mean that none of the individuals showed a native-like ERP response. Rather, a null effect might be based on opposite effects (a positive going effect in one set of individuals and a negative going effect in others) canceling each other out. In a similar way, biphasic responses can be a spurious result of averaging (Osterhout, 1997; Nieuwland and Van Berkum, 2008; Tanner and Van Hell, 2014; Tanner et al., 2014). Before we draw any strong conclusion that a group of learners' processing of gender agreement qualitatively differs from natives, it is important to identify varying patterns in each of the groups. Furthermore, there may be predictors of native-likeness in L2 learners, such as age of acquisition, proficiency, language exposure and use, that may explain variance within the group (e.g., Weber-Fox and Neville, 1996, 1999; Rossi et al., 2006; Steinhauer et al., 2009; Tanner et al., 2014). Understanding which individual difference factors, if any, are associated with the outcome in L2 learning is a fundamental question which is difficult to answer with group-based analyses, and might also help us determine the source of some of the mixed patterns of results in L2 gender research.

Using the P600 as a measure of native-likeness, we tested whether sufficiently proficient late L2 learners can show native-like syntactic processing, even if (1) gender marking in the L1 is implemented differently and (2) the L2 gender system is opaque. We investigated the ERP responses of native speakers and Romance learners of Dutch to anomalies in constructions that are relatively easy to acquire (i.e., non-finite verbs) and those that have been shown to be more difficult (i.e., gender). In addition, we varied the modality in which the stimuli were presented, in order to investigate whether visual presentation might contribute to the lack of sensitivity to gender in the Romance group reported in previous research (Sabourin and Stowe, 2008). The non-finite verb violations elicited a biphasic N400-P600 effect in both native speakers and second language learners. However, in contrast to the native speakers, the learners only showed evidence of an N400 in the auditory and not the visual condition, although the statistical support for this difference is weak³. Also, the amplitude of the N400 effect was somewhat smaller than in the natives. For the gender violations, we found a clear P600 in natives, but not in L2 learners.

The effects of modality were quite subtle. We had hypothesized that increased processing demands in the visual modality might interfere with immersed learners' responses to grammatical violations and that they might show more native-like responses in the auditory modality. This hypothesis receives some support; the modulation of the N400 effect in non-finite verb violations in learners was in the hypothesized direction, with a native-like effect in the auditory but not the visual modality. However, for gender agreement learners failed to show sensitivity, regardless of the modality. Thus, the suggestion that the difference between Loerts' 2012) results for Polish speakers on the one hand, and Sabourin and Stowe's (2008) results and our current results for Romance speakers on the other, cannot be attributed to the difference in modalities.

In contrast to the modality effects, violation effects and group differences therein were robust. Before accepting the group patterns, it is important to examine the role of individual differences. A biphasic pattern may reflect the summation of single effects originating in two different groups of participants (Osterhout, 1997; Nieuwland and Van Berkum, 2008; Tanner and Van Hell, 2014; Tanner et al., 2014). Even more crucial for the current experiment, the absence of an effect in the L2 group may be due to variability, with some individuals showing the pattern found in native speakers, while others show no effect or even an opposing effect (Foucart and Frenck-Mestre, 2011). Inspection of individual differences for the gender violations confirmed that the grand average ERP patterns we report are representative of the majority of the individuals in each group. In contrast to natives, who consistently showed large P600 effects (Figure 6, bottom left panel), learners consistently failed to demonstrate any form of sensitivity to gender violations (Figure 6, bottom right panel). This result was confirmed by the fact that none of the participant characteristics we tested (increased proficiency or gender knowledge, earlier age of acquisition, longer length of residence or high percentage L2 use) was associated with a larger P600. In this sense, the current experiment replicates the pattern found by Sabourin and Stowe (2008); even highly proficient Romance learners of Dutch appear to have persistent difficulties in learning to use Dutch gender.

Turning to the non-finite verb violations, examination of the native speakers confirms that the biphasic pattern N400/P600 seen in this group is present in the majority of the individual participants (see Figure 6, top left panel). This biphasic effect in response to non-finite verb violations in natives has been found before (Kutas and Hillyard, 1983; Sabourin and Stowe, 2008; Loerts, 2012). As can be seen in Figure 6 (top right panel), many learners' responses were within the native range, showing evidence of the biphasic pattern, although this is primarily evident for the auditorily presented materials. Some individuals are less native-like; for this structure the P600 effect magnitude in the L2 group was found to be modulated by the percentage of use of the L2 in daily life. Use is not the only important factor for native-like attainment of syntax processing however; even the learners with the highest amount of daily practice in an immersed setting still show persistent problems with gender agreement.

Despite their failure to show native-like gender processing, the evidence suggests that the Romance learners are highly proficient. In addition to the off-line measures of proficiency (C-test and gender assignment) and online accuracy at ungrammaticality detection, which are within native range for a number of the participants, the evidence from the biphasic N400-P600 pattern provides a strong argument for high proficiency. Finding early ERP effects in response to grammatical violations like the N400 seen here is unusual in L2 research. Although both Loerts (2012) and Sabourin and Stowe (2008) found evidence of a biphasic pattern for their native groups, neither found the N400 in their L2 learner groups. According to Steinhauer et al. (2009), biphasic patterns are one of the latest stages of morpho-syntactic proficiency in late L2 acquisition. The fact that our learners were able to reach this stage for non-finite verb agreement, but that they cannot get past the initial stage of not showing any brain response differences for correct vs. incorrect use of gender agreement provides strong support for the difficulty of the acquisition of this element in Dutch L2 acquisition. This highlights the complexity of acquisition of the Dutch gender system, even by learners with a gender system in their L1. Furthermore, it emphasizes the fact that language learning aptitude is not an all or none phenomenon, but may vary widely between constructions.

Our results further illustrate the large discrepancy between online and offline processing measures in L2 acquisition research. Both the behavioral results of the gender assignment task and the sentence-final grammaticality judgments during the ERP recordings for gender violations indicate moderate to good knowledge of Dutch grammatical gender in the learner group. Yet, we observed a complete lack of response to these violations in the ERP signal. This reveals a discrepancy between offline knowledge of grammatical gender concord and the use of agreement knowledge during online processing. The lack of a significant relation between the magnitude of the P600 responses to gender violations and the score on the gender assignment task rules out the possibility that only learners with better offline performance are able to show online effects. The behavioral difference between the visual and the auditory modality, with performance being slightly worse for grammaticality judgments in the auditory modality, was also not reflected in the ERP signal for gender violations. These results illustrate that second language learners can develop successful strategies to cope with gender processing difficulties. These alternative routes, however, apparently take more time and are qualitatively different from what we observe in online native processing.

The results of the current study leave us with a puzzle; why do Romance learners of Dutch show such persistent problems with gender processing? Our results confirm that gender is difficult to process for late Romance learners of Dutch, compared with the results of studies targeting other languages. We replicated Sabourin and Stowe's (2008) findings, in the sense that our Romance learners likewise did not show native-like responses to gender violations, regardless of modality, although they showed responses to non-finite verbs that were close to the native model⁴. The factors most commonly suggested in the literature as to why gender or other forms of grammatical processing might be problematic do not appear to explain these results. Proficiency clearly plays some role in native-likeness in general (Steinhauer et al., 2006; McLaughlin et al., 2010), but as we argue above, our learners were quite proficient, certainly comparable to those in other studies in which learners have shown P600 effects for gender (Tokowicz and MacWhinney, 2005; Frenck-Mestre et al., 2009; Gillon Dowens et al., 2010, 2011; Foucart and Frenck-Mestre, 2011, 2012; Loerts, 2012). Also, our proficiency measure does not correlate with the magnitude of the ERPs.

Other potential explanatory factors involve the language experience of the learner, such as age of acquisition (Weber-Fox and Neville, 1996; Kotz et al., 2008) and exposure to and use of the L2 (Gardner et al., 1997; Flege et al., 1999; Dörnyei, 2005; Tanner et al., 2014). It is true that the studies reported by Frenck-Mestre and colleagues have generally tested earlier learners (with onset of acquisition in their teens rather than twenties and later). However, other studies have demonstrated native-like gender processing even for relatively late learners (Tokowicz and MacWhinney, 2005; Gillon Dowens et al., 2010). Furthermore, in the current study we did not even find a trend toward better performance for younger learners, making it again unlikely that this is the (only) decisive factor for native-likeness. The amount of L2 use also failed to explain the failure of the Romance learners to show online sensitivity to gender, even though, as our own results show, this can be important for native-likeness for other aspects of grammatical processing, like verb agreement. Length of residence, which impacts overall exposure, also showed no correlation with sensitivity to gender.

Failure to achieve native-like processing has also been linked to dissimilarity between L1 and L2 (Tokowicz and MacWhinney, 2005; Sabourin and Stowe, 2008; Foucart and Frenck-Mestre, 2011), as well as characteristics of the target language (Sabourin and Stowe, 2008; Loerts, 2012). Following this line of argumentation, Dutch and Romance languages may simply be too different from each other, which, combined with the fact that the Dutch gender system is relatively opaque, results in a very difficult challenge for native-like attainment. The lack of transparency of the Dutch gender system might explain why our Romance learners failed to show native-like processing for this characteristic of the language, as opposed to the much more transparent non-finite verb manipulation. For gender, previous research has shown that native-like processing is possible even in constructions with competition from an L1 gender system when a relatively transparent target gender system is to be acquired in L2 (Frenck-Mestre et al., 2009; Foucart and Frenck-Mestre, 2011; Gillon Dowens et al., 2011). In contrast, Loerts' study suggests that an opaque system is more difficult to acquire, since only her most proficient learners are able to show P600 effects, which are additionally somewhat smaller in amplitude compared to the natives. It remains an open question as to why, in contrast to Loerts (2012), even the most proficient learners in the current study did not show a P600. More research is needed to determine whether characteristics of the L1 or other (confounding) factors are at play in determining which individuals overcome the challenge of an opaque gender system.

One final point we would like to make is that, although we did not find extensive effects of stimulus modality, this factor is nevertheless of importance. As we noted, the early responses to ungrammaticality like the N400 in the biphasic response seen here are not generally found in late L2 learners, which has been taken as a sign of lack of native-likeness. It is possible that they have been missed due to the use of visual materials, since this effect was only seen in the auditory modality. Although we saw no effects on the amplitude of the P600 effect, certain populations may be affected more than others. Learners who do not share the same writing system in their L1 and L2, for instance, might have more difficulty automatizing their usage of the new alphabet (Koda, 1999; Wang et al., 2003). For these learners, the use of auditory materials might be a crucial prerequisite to obtain an accurate measure of their abilities. On the other hand, those whose learning has taken place with an emphasis on written materials may show less response when auditory materials are used. Given the large diversity of L2 speaker populations with respect to typological distance (both with respect to grammar and writing systems) and type of learning environment (immersion vs. classroom), it is important to be aware that the testing modality might influence the results, both in offline and online tests.

In conclusion, we can say that online grammatical gender processing is particularly difficult for Romance learners of Dutch, even at high levels of proficiency and with large amounts of L2 exposure and use in a natural setting, and regardless of testing modality. In contrast, responses highly similar to the native model are possible for a more regular and transparent structure (non-finite verbs), for which responses are modulated by both testing modality and L2 use. In contrast, the problems with gender are persistent and not affected by these factors, demonstrating the complexity of (late) L2 acquisition of the opaque Dutch gender system.