Reaction time experiments such as the word naming task are widely used to explore differences between L1 and L2 language processing. However, there continues to be debate about how to interpret naming latencies. It appears to be a major challenge to relate any of the many factors that are involved in the naming of a word (word frequency, word length, number of syllables, size of the neighborhood, the transparency of matching graphemes to phonemes, the pronunciation, and the degree to which word meaning plays a role) to a particular portion of the total naming time (de Groot et al., 2002). Even though the same factors can be assumed to be involved in naming a word on different occasions, the naming latency is subject to substantial variability, even when repeating the same word in succession (Kello et al., 2008) or when recording the naming latency of the same words on many different occasions (de Bot and Lowie, 2010). Since the components involved in the naming process flexibly, interactively and continuously contribute to a naming latency, the resulting naming latency is never an absolute constant.

Therefore, keeping in mind that the contribution of all the different factors involved in naming will be variable across tasks, languages, and participants, the current study aims at clarifying the degree to which automatic and controlled processing play a role in naming from an interaction-dominant perspective. Rather than trying to isolate the factors of interest in L1 and L2 naming, automatic grapheme-to-phoneme conversion and semantic involvement, the current study will use different tasks in which the relative contribution of these processes is assumed to be different. Analyzing the variability of naming latencies (RTs) during the task rather than the overall time it takes to name the words presents the opportunity to look at the relative contribution of automatic (grapheme-to-phoneme conversion) processes and more conscious, attention-demanding processes (semantic involvement) in L1 (Dutch) and L2 (English) word naming. With the possible exception of the most advanced speakers of an L2, word naming results consistently show language performance to be slower in an L2; however, longer naming latencies in the L2 do not necessarily indicate L2 naming to be qualitatively different from L1 naming (Segalowitz et al., 1999). Focusing on the variability of reaction times across a word naming task rather than an overall mean outcome can provide insight into the naming process as it is unfolding, and thus, into the nature of language processing.

A possible explanation why L2 language performance is usually slower in reaction time experiments could be simply the degree of use and practice. The underlying processing may be qualitatively similar in L1 and L2 language use in advanced L2 speakers, but due to less practice in using the language these processes could be less automatized and thus slower. However, another explanation could be that L1 and L2 language processing are qualitatively different, even for advanced L2 speakers. Using a language successfully requires the interaction of several components, such as mapping graphemes to phonemes, articulating sounds, and semantics. The relative contribution of these different components could be different between languages depending on proficiency and frequency of use. Looking at the speed of processing alone is not enough to distinguish between these accounts.

The central aim of the current study is to disclose the underlying processing involved in naming words in the L1 and L2, and to see if these processes are qualitatively different. A secondary question is if any differences that are found in the underlying processing, are attributable to the languages being a speaker’s L1 or L2, or are to do with differences between the languages. In order to answer these questions, the focus will be on the degree of automatization in naming words in an L1 and L2 as is shown from patterns of variability, and on assessing the relative contribution of semantic involvement in the naming process.

Since automatization and semantic involvement are part of language processing and take place over time, this study will use non-linear analyses to shed light on the underlying processes involved in naming. Originating in biology and ecology, but now also used extensively in other fields of cognitive science, spectral analysis is used to look at variability in response time data that has traditionally been regarded as irrelevant ‘noise’ but has been found to reveal the degree of automatization of mental processes (Van Orden et al., 2003; Kello et al., 2008; Wijnants et al., 2009; Lowie et al., 2014). Looking at patterns of variability in a word naming task in L1 and L2 allows us to look beyond the speed of processing at the difference in optimal coordination using the mother tongue or a second language.

The word naming task is a quick and undemanding task suitable for gathering language production data of multilingual speakers, without the drawback of confounding lexical processing with discrimination, as could be the case for the lexical decision task (Balota and Chumbley, 1984). For the purposes of the present article word naming also has the advantage of being continuous and can be applied without inter-experimental conditions, which allows for analyzing the set without interruptions and looking at processing over time. However, there is some debate concerning what processing the naming task actually taps into. Word naming is driven by orthography alone, in contrast to for instance picture naming, which is necessarily conceptually driven. In word naming, fast pronunciation is usually the only instruction; in this sense, semantic involvement could be seen as redundant with regards to the task goal. It is clear that semantic involvement is not a necessary requirement for pronunciation; one can pronounce a non-word or a word in a language unknown to the reader. Many dual route theorists indeed assume skilled readers to not use the semantic pathway when naming single words (Shibahara et al., 2003); however, semantic involvement has been found to play a role in naming low frequency, irregular words (de Groot et al., 2002; Shibahara et al., 2003). It thus remains unclear if, and to what degree, meaning plays a role in the pronunciation of single words. The amount of semantic involvement has also been found to vary between languages due to orthographic depth; that is, the transparency of a languages correspondence between phonology and orthography. The more complex a language’s spelling-to-sound conversions, the greater the orthographic depth (Katz and Feldman, 1983; Shibahara et al., 2003). There is some indication that for the multilingual speaker, reliance on semantics in reading varies for different languages with different orthographic depth, since more complex and inconsistent spelling-to-sound conversion rules may require semantic information in order to select a correct pronunciation (de Groot et al., 2002).

A word naming task is a suitable method for obtaining long, continuous data series in either the L1 or L2, and thus provides the opportunity to look at language processing over time. The different components (production and understanding) that contribute to natural language use are present, but a word naming task draws heavily on the production part of language use. In regular word naming, an automatic orthography-to-pronunciation link could be used in which semantics plays a minor role (de Groot, 2011, pp. 157–158). Due to a lot of practice, word pronunciation may be more automatized in a first than a second language. In order to test whether the underlying lexical processing is also more automatized in L1 naming than in L2 naming, a semantic categorization version of the naming task was devised, to ensure semantic processing during naming. The different versions of the naming task were tested on two different groups of participants; each participant was tested in the L1 and L2.

The resulting reaction time series of the word naming experiments were analyzed using spectral analysis, so as to gauge what happens over time when using the L1 or L2, rather than look at the product or result of the task. Spectral analysis can be used to look at the variability in the data, and gives an estimate of how random or structured the variability is. A structured pattern of variability is linked to more automaticity and a more optimal coordination of underlying processes. Of interest was the degree of automaticity in L1 and L2 language processing, and the difference between the standard word naming and the semantic categorization naming to assess the degree of semantic involvement in L1 and L2 language processing.

A consistent finding when comparing L1 to L2 naming is a speed difference; naming in the L2 has repeatedly been shown to be slower than naming in the L1 (e.g., experiment 3, Kroll and Stewart, 1994; La Heij et al., 1996; Van Wijnendaele and Brysbaert, 2002), even for fluent, (nearly) balanced bilinguals (Potter et al., 1984). These speed differences in L1 and L2 naming have been linked to automatic spelling–sound conversion and semantic involvement in naming, as in the 2002 study by Wijnendaele and Brysbaert. They found longer naming latencies for the L2 to be modulated by frequency; the difference in RTs was larger for low frequency words than for high frequency words, and more pronounced in L2 than in L1. The frequency effect is considered a marker of lexical involvement in the naming process, which leads the authors to conclude that the absence of a strong frequency effect in L1 naming suggests that L1 naming is largely mediated by automatized, non-lexical spelling–sound conversions, whereas the clear frequency effect found for naming in the L2 is taken to indicate that correct word naming in the L2 requires more lexical (whole word) mediation than word naming in L1, on account of there being a less powerful non-lexical route available for the L2 (Van Wijnendaele and Brysbaert, 2002). ERP studies, which are very sensitive to the time course of processing, seem to suggest that speed differences do not arise from differences in semantic processing. Several studies using event-related potentials (ERPs) have been conducted to look at differences in L1 and L2 processing. Even though these studies have found significant differences in syntactic L1 and L2 processing, semantic anomaly detection in sentence processing in high proficient L2 learners yielded very similar ERPs (Hahne, 2001; Hahne and Friederici, 2001) or a small delay of around 40 ms in L2 processing compared to L1 processing (Ardal et al., 1990).

de Groot et al. (2002) try to disentangle recognition and production as the locus of the effect of 18 variables. They compare lexical decision, standard word naming and delayed word naming to look at the “functional overlap” (Grainger and Jacobs, 1996) between the tasks and to see if a selection of 18 variables including word length, neighborhood, meaning, onset, frequency, etc. affected the processing of Dutch and English words in Dutch–English bilinguals differently. One of the most salient effects found by de Groot et al. (2002) is that standard naming and lexical decision responded very differently to the word manipulations, while the two tasks produced similar results in the L2 English. One of the language effects found by de Groot et al. (2002) most relevant to this study, is a semantic effect found for English naming but not for Dutch naming.

However, there is always a possibility that differences found between L1 and L2 language processing are confounded with differences that are actually attributable to differences between the languages themselves. Different languages may call for different processing strategies. An relevant example in the current context is the orthographic depth hypothesis (Katz and Feldman, 1983). The orthographic depth hypothesis concerns the differences between languages in the degree of complexity of their spelling–sound conversions. In the Katz and Feldman (1983) study, Serbo-Croatian and English naming are compared. They use Serbo-Croatian as an example of a language with shallow orthography; that is, very simple and direct spelling-to-sound correspondence. English, in contrast, has a deep orthography, with a complex, abstract correspondence between orthography and phonology. This leads the authors to hypothesize that readers of a shallower orthography may depend less on semantic mediation, and instead rely on fast, simplified spelling-to-sound correspondences. The absence of a semantic priming effect in Serbo-Croatian naming but not in English naming leads the authors to conclude that this is indeed the case. This effect was also found by Cuetos and Barbón (2006), who found semantic variables to affect Spanish word naming times less than orthographically deeper languages. For the present study, the orthographic depth hypothesis may also be relevant in explaining differences in Dutch and English naming; Dutch, like Serbo-Croatian and Spanish, has a shallow orthography with a more direct spelling-to-sound correspondence than English (van den Bosch et al., 1994). It is thus conceivable that in the naming of Dutch words, there also may be more of a reliance on fast and automatic spelling-to-sound conversion and less on semantic involvement.

Some research indicates at least reduced semantic involvement in word naming as compared to picture naming, lexical decision or translation (Lupker, 1984; Potter et al., 1984; Kroll and Stewart, 1994; de Groot et al., 2002). For instance, the poor recall of words in word naming as compared to picture naming has been interpreted as signaling reduced conceptual mediation and reduced depth of processing (Smith and Magee, 1980; Potter et al., 1984). de Groot et al. (2002) show that frequency effects are smaller in word naming than in lexical decision tasks, which is taken to be a sign of reduced involvement of semantic representations (de Groot et al., 2002). de Groot (2011, pp. 157–158) points out that for languages with an alphabetic script and regular grapheme–phoneme relations, the responses can be simple script-to-sound conversions, without any semantic processing. This would account for the difference found for the time it takes to name written words and pictures. It has been known for over a century that written words are named 200–300 ms faster than pictures (Cattell, 1886; Potter et al., 1984; Ries et al., 2015). The explanation for this difference in naming latencies is that a participant can start pronouncing a written word without the necessity of understanding it or retrieving the underlying concept. In this view, the process of naming consists of a series of stages that one has to go through to ultimately deliver the spoken word. In the case of naming a picture, the assumption is that the picture’s underlying concept first has to be found. This will make the identification of the correct word associated with this concept possible, and only then pronunciation can start. For written word naming, with fast pronunciation being the only goal, retrieving the underlying concept would be an essentially redundant stage that would only slow down performance, and pronunciation can start right away. An investigation into the time course of semantic and phonological encoding in a picture naming and a listening task using ERPs confirms the different order of semantic and phonological processing in different tasks and circumstances. A go/no go decision task based on phonological (first letter) or conceptual features (animal or object) showed that in picture naming, semantic processing preceded phonological processing by 170 ms. In listening, the effect was reversed, albeit smaller; phonological processing preceded semantic processing by 85 ms (Rodriguez-Fornells, 2002).

Dual route models of word recognition visualize the stages from print to spoken word [dual-route cascaded (DRC) model (Coltheart et al., 2001) or the connectionist dual process (CDP+) model of reading aloud (Perry et al., 2007)]. Figure 1 shows the DRC model (Coltheart et al., 2001), which is used here to visualize the processes that are potentially involved when naming a written word. This model depicts visual word recognition to be the product of the combined activity of an orthographic lexicon (the lexical route) and a rule-based grapheme–phoneme conversion (GPC) system (the non-lexical route). Word recognition in dual route models is based on visual encoding of orthographic information. The GPC system is necessary to account for the pronunciation of non-words or unknown words. Reading of (known) words makes use of both the lexical and non-lexical route; the difference being that through the lexical route, the pronunciation of the word is retrieved whole. After a lexical entry has been located in the orthographic lexicon, the matching phonology is accessed in the phonological lexicon. This information is then forwarded to the phoneme system, where the pronunciations of both routes comes together and the correct pronunciation can be chosen. Only use of the lexical route will result in the correct pronunciation of words with irregular grapheme-to-phoneme correspondence. Through the lexicon, the semantic system can be accessed. However, there are direct links between the lexicon and the phoneme system, which means that involvement of the semantic system is an optional ‘detour.’ Due to their modularity, dual route models (and most schematic models of language processing) feed into a dichotomous perspective on language processing; there either is semantic involvement, or there is not.

A possible alternative to dual models of word recognition is offered by parallel distributed processing (PDP) models, such as the PDP triangle model by Harm and Seidenberg (2004). The triangle model is a computational, connectionist model that aims to be biologically plausible, with a built-in ability to learn and self-organize. The 2004 model was focused on the access of word semantics directly through orthography, or indirectly, via phonology. This last way was envisaged as taking more time. Also in dual models of word recognition, when processing takes place through the lexical route, phonological information comes in relatively late, after the lexical item has been selected. Strong phonological models of visual word recognition make no distinction between a lexical and a non-lexical route; these models envisage phonology and semantics playing a major role in visual word recognition from the onset (Frost, 1998; Carreiras et al., 2014). In Frost’s computational model of word naming, phonology is always prelexically assembled and, if the task requires detailed phonology, lexically supported (lexical representations are needed, for instance, for the fully articulated, correct pronunciation of words with ambiguous spellings) (Frost, 1998). The model thus proposed by Frost is not dichotomous, but continuous, with both prelexical assembly and lexical/semantic involvement always playing a role in naming. This is in line with a study using a backward masking paradigm by Berent and Perfetti (1995), who find that the phonological representation of English CVC words is computed in two processing cycles with different time courses. In the first cycle, which is a fast and automatic prelexical process, consonants are computed. This generates a phonological skeleton of consonantal information. The second cycle is slower and involves attention-demanding processing; lexical information is used to fill in the vowels, which are the main source of lexical ambiguity (Berent and Perfetti, 1995). As Frost points out, if both lexical and prelexical computations affect the naming task, the focus of investigation should not be on whether there is lexical involvement on naming, but rather on the relative use of these processes in naming.

In models such as the ones discussed above, semantic representations are one of many components of stored lexical information. Lexical access in this view entails finding a simplified orthographic/phonological representation, whereas access to meaning is a later stage in word processing that can occur only after the lexical representation has been retrieved. The dissociation between lexical and semantic representations has long been taken as a necessary distinction to account for fast word processing. However, in the past decade this idea has been challenged. Elman points to the problem of where to draw the boundary between the lexicon and other types of linguistic knowledge. He found that many contextual factors (agent, tense, patient, location, filler of the instrument role, information given in the broader discourse context) influence expectations regarding the argument a verb will take. It would not do to list separate entries for every individual occurrence of a word, but any decision on what to include in the lexicon would also be rather arbitrary. He suggests “words do not have meaning, they are cues to meaning” (Elman, 2004, 2009, 2011). Spivey (2007) also contests the idea of discrete, static representations, and stresses the fluid nature of cognitive processing, claiming that mental (whether lexical, semantic, or otherwise) representations should be thought of as processes: “sparsely distributed patterns of neural activation that change non-linearly over the course of several hundred milliseconds, and then blend right into the next one” (Spivey, 2007). Applying these findings to the current study, it would follow that it is not necessarily a fruitful path to follow to think of a sequence of events as pre-lexical (e.g., phonological assembly) or post-lexical (e.g., semantic access), or to think of semantics and meaning as being mediated by lexical entries in a mental lexicon, nor would it be interesting to see whether or not semantics plays a role in word naming. Assuming the activation of meaning to be a gradual process that plays a role in word recognition, but perhaps to a different degree depending on for instance language dominance, the focus of interest would then be the relative contribution of meaning activation in naming.

Models such as the dual route models of word recognition do not specify if and how the word recognition process would be modulated for multilingual language use. A lot of research has focused on differences between L1 and L2 language processing, such as the influence of the L1 on L2 processing, but as Lemhöfer et al. (2008) point out, these studies have been conducted with carefully selected and few test items, with the aim of uncovering differences. Whether or not these differences would still exert an influence in natural language processing is uncertain; after all, in less controlled conditions it may be balanced out by many other factors playing a role in processing. Lemhöfer et al. (2008) conducted a study of bilinguals with different L1’s (French, German, and Dutch) to look into the L1 specific influence on L2 processing. They used a multiple regression model that allowed them to look at the influence of many variables, including word frequency and length, concreteness and meaningfulness and many others. Apart from cognates, L1 specific influence on L2 processing was absent, and they conclude that L2 processing is “language driven,” meaning that within-language factors seem to determine the adopted processing strategies. However, even though bilinguals with different L1 backgrounds process the L2 in the same way, they did find differences in the language processing of native speakers and L2 speakers, especially on word variables related to frequency and ways of occurrence. These variables exert a much greater influence on L2 processing than on L1 processing, and the authors conclude that L2 processing is thus “fundamentally different” from L1 word processing (Lemhöfer et al., 2008, p. 27).

That L1 and L2 language processing are different, has been made visible by a range of neuroimaging studies. A marked difference is the level of attention or control needed for bilingual language processing. In an fMRI study looking at age of acquisition and proficiency effects in L1 and L2 processing, Wartenburger et al. (2003) found more extensive cerebral activation during semantic judgment tasks for L2 processing than L1 processing. In a set of studies using EEG and fMRI, Rodriguez-Fornells et al. (2002) compare monolingual and bilingual language processing and found more left prefrontal activation in the bilinguals, associated with increased control and language inhibition. Important factors that modulate differences in bilingual versus monolingual language production are relative proficiency in each language, as well as the level of competition between both languages (Abutalebi and Green, 2007; Kroll et al., 2008; Green and Abutalebi, 2013). Meschyan and Hernandez (2006) show in an fMRI study differential processing in languages that differ in orthographic depth; Spanish is a very transparent language, and English has a deeper orthography. They found that different brain regions were involved when English–Spanish bilinguals were reading Spanish words than when reading English words. The study shows the bilingual adapting to the most optimal lexical access strategies of the target language; a skill that is likely to develop with increasing proficiency (Meschyan and Hernandez, 2006). In an overview study looking at results of a large number of fMRI and PET studies that investigate bilingual language production, Abutalebi and Green (2007) argue that an increase in proficiency is accompanied by a shift from controlled to more automatic processing, and by a reduction in prefrontal activity. Increasing proficiency will lead to a dynamic adaptation of the language system and the control networks. However, the authors point out that it is likely that the bilingual will not process the L2 identical to the way a monolingual speaker of that language would. In the same vein, they hypothesize the bilingual’s processing of the L1 is unlikely to be identical to a monolingual’s processing of the L1.

The present study aimed to investigate the relative contribution of automatic processes (grapheme-to-phoneme conversion) and attention-demanding processes (semantic involvement) in L1 and L2 language production using a standard word naming task and one that required semantic categorization. Spectral analysis was used to look at the degree to which the L1 and L2 language subsystems were optimally coordinated and automatized during word naming, and the degree to which semantic processing played a role in word naming.

A fractal pattern of variability was found for both L1 Dutch and L2 English word naming. The spectral slopes found for L1 Dutch standard word naming (-0.38) were significantly steeper than the spectral slopes found for the English standard naming condition (-0.19). It is important to stress that the steepness of the spectral slopes is not related to the speed of processing; even though L2 naming is consistently slower, this does not lead to a relative difference in the steepness of spectral slopes. Speed of processing and scaling are independent of one another. As indicated, a fractal pattern of variability is associated with optimal coordination and automatization of the mental subsystems contributing to a task. The fractal structure in the L2 data shows a more decorrelated pattern closer to random variation. The stronger fractal pattern in the L1 data indicates a more optimal coordination of mental subsystems when naming L1 words than when naming L2 words. In other words, the appearance of a string of letters on the screen that correspond to an L1 word leads to a highly automatized, smooth response that is consistent with an account of automatic spelling-to-sound conversion in the L1. The same degree of automatization does not seem available in the L2.

The same naming task in the semantic condition was conducted to ensure that the processing that was being investigated was not only governed by automatic spelling-to-sound-conversion, but also semantic processing. Comparing this task to the standard word naming allows us to infer the degree of semantic processing present in standard word naming; if no difference between the conditions were found, the same degree of semantic processing must already be present in standard word naming. However, the semantic categorization condition did have a differentiating effect in the L1 and L2; a comparison of the L1 standard naming condition with the L1 semantic naming condition shows a more decorrelated fractal pattern of variability in the L1 semantic naming condition (-0.38 for L1 standard naming and -0.28 for L1 semantic naming). The same effect was not found in L2 naming; both standard L2 naming and L2 semantic naming generated the same mean spectral slope of -0.19. This finding is in line with the observation that L1 naming relies more than L2 naming on automatic spelling-to-sound conversion. The semantic categorization condition would have forced more attention-demanding lexical involvement than regularly used in the reading aloud of L1 words, leading to a variability pattern that was less fractal than the one observed for L1 standard naming. If automatic spelling-to-sound conversion is less available in the L2, as suggested by de Groot et al. (2002) then semantic/lexical involvement already plays a significant part in L2 naming, and the manipulation would not have made a difference to the normal procedure of reading aloud single L2 words.

To find out whether the observed difference in slope statistics comes down to a difference between L1 and L2 processing, or between Dutch and English language processing (recall the orthographic depth hypothesis) it is useful to compare the languages in the semantic categorization condition, and to also compare these results to L1 English standard naming, as obtained by Van Orden et al. (2003)¹. For easier reference, the Figure 5 combines the results obtained in the present study, and those found by Van Orden et al. (2003).

The spectral slope found for the Dutch semantic condition was less steep than the slope found in the standard naming condition, but at -0.28, it was still steeper than the spectral slope of -0.19 found for the English semantic condition. So, even if the degree of lexical involvement is manipulated for both languages, a language effect remains intact, indicating more automatized processing in the L1. However, these results do support the orthographic depth hypothesis, which classifies English as having a deep orthography with an inconsistent spelling-to-sound correspondence that requires lexical mediation for correct pronunciation. The slope of -0.28 found for the L1 Dutch semantic condition corresponds almost exactly to the mean spectral slope of -0.29 reported by Van Orden et al. (2003) for English L1 standard word naming (Van Orden et al., 2003). The differences in the relative contribution of automatic processes (grapheme-to-phoneme conversion) and attention-demanding processes (semantic involvement) in L1 and L2 language use found in this study seem to be attributable to two different causes. The steeper slopes in L1 standard Dutch naming (-0.38) than in L1 standard English naming (Van Orden et al., 2003) can then be attributed to the difference in orthographic depth of the languages; Dutch having the shallower orthography, allows for more automatic spelling-to-sound conversion. However, the L2 status of English for the participants in this study did still generate an effect beyond the deeper orthography, participants using more attention-demanding processing in the L2 than the English native speakers in the Van Orden et al. (2003) study, as reflected in spectral slopes of -0.19.

This leads to another observation on account of the advanced L2 learners participating in the current study. Even though they do not achieve the optimal coordination of underlying subsystems when using the L2, they are adopting language processing strategies with a higher relative contribution of lexical mediation than is usual in their L1, thus adopting native-like language processing strategies, albeit less optimally organized than native speakers.

The goal of this investigation was to look at the relative contribution of automatic grapheme to phoneme conversion processes and more conscious, attention-demanding processes (lexical/semantic involvement) in L1 Dutch and L2 English word naming. As hypothesized, the relative contribution of automatic vs. attention-demanding processes in the L1 and L2 was not equivalent, even though the learners participating in this study were near-native L2 speakers. The scaling relation found for L1 naming as compared to L2 naming was closer to the scaling relation indicating automatic, optimal coordination of mental processes, probably achieved through more regular use and practice of the L1. This conclusion is in line with Lowie et al. (2014), who looked at a single native Dutch speaker who was very proficient in the L2 and regularly spent time in the L2 environment; they found steeper spectral slopes for the language that was recently used, regardless of it being the L1 or L2, and a decorrelated fractal structure for the language that was not recently used.

In addition to, and related to the relative contribution of automatic spelling-to-sound conversion, the focus of interest was the relative contribution of more attention-demanding, semantic processing. The semantic condition was found to affect patterns of variability in the L1 only, leading to the conclusion that L2 speakers rely more on lexical involvement in naming words. Even when forcing more contribution of semantic processing in the L1, the spectral slopes still indicated a higher contribution of automatic spelling-to-sound conversion processes in L1 naming than in L2 naming. We would therefore like to propose an amendment to the conclusion reached by Lemhöfer et al. (2008), who conclude that L2 processing is mainly language driven. The present findings with a group of very proficient bilinguals, would lead us to the conclusion that it is in part language driven, and determined by the language properties such as orthography of the target language. However, language use also is also user driven, determined by the bilingual language system, including factors such as frequency of use that constrain the language production process and lead to more controlled processing.

This study is not the first to find L2 processing is less automatized than L1 processing (Rodriguez-Fornells et al., 2002; Wartenburger et al., 2003; Abutalebi and Green, 2007; Green and Abutalebi, 2013). However, by looking at the relative contribution of automatic and controlled processing from an interaction-dominant perspective, it offers an explanation how and why this might be the case. From an interaction-dominant perspective, in which there is a coupling and coordination of contributing subsystems, different languages within a multilingual speaker being more or less automatized becomes explicable, and it becomes possible to examine exactly which processes contribute to making the coordination of underlying subsystems more automatic or controlled, in this case by manipulating the degree of semantic involvement in the task. The dynamic approach used here shows a quantifiable, testable way of looking into the self-organizing behavior of the language system, and how it constantly adapts, both to language specific factors such as a language’s orthographic properties, and to the language user’s capabilities and constraints related to the frequency of use of a language.

This differential degree of automatized and controlled processing in L1 and L2 naming has implications for how to interpret the extensive body of research into multilingualism that uses this task to look at word processing differences between languages in multilinguals. Earlier research already indicated care should be taken in interpreting evidence on word recognition that may be confounded with task-specific processing (Balota and Chumbley, 1990; de Groot et al., 2002), since for instance a lexical decision task may tap into different processing strategies than a word naming task. However, the results from the current study show that even using the same task in different languages may tap into different processing strategies. Related to this, and pointed out by Abutalebi and Green (2007), is that when the system adapts toward strategies that are most appropriate to process the L2, the entire system changes. Therefore, increasing proficiency in optimally processing the L2, and specifically L2 learners using semantic processing more in the L2 than they are used to in the L1, could change the way in which the L1 is processed as well. For the current study, this means that it would be interesting to compare the results to monolingual Dutch speakers as well, since the participants in the current study possess a language system adapted to processing an L2. Furthermore, whereas the current study shows semantic reliance to differ between languages within an individual, different individuals also differ in the degree of semantic involvement used in word naming (Woollams et al., 2016).

These results taken together indicate caution is needed when interpreting studies that are generalized over individuals, monolinguals and bilinguals, languages or tasks. If word naming in the L2 tests ‘deeper’ processing than naming in the L1, caution is needed when comparing the results of multilingual naming. When a task such as word naming is used in bilingual research, the aim is to be able to generalize about a variable that has been manipulated. The assumption is that apart from this manipulated variable, the same task in different languages is testing the same underlying processes. The (multilingual) individual’s adaptability to task, circumstance and environment is underestimated in such direct comparisons; this study suggests caution is needed when comparing cross-linguistic results of word naming tasks.

The study was exempt from ethical considerations. The experiment concerned a non-invasive reaction time experiment. Participants were adult, healthy university students, who participated voluntarily and received a small reward. For some participants, participation was part of course credit.

RP, principal author, carried out the study, analyzed the results, and wrote most of the paper. WL, promotor and daily advisor, has been involved in designing the study and providing advice and feedback along the way. KdB, co-promotor, was involved in the design of the study and giving advice and also offering feedback during the writing of the paper.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.