Highlights
-
• First study to explore simultaneous bilinguals’ phonotactic knowledge and processing.
-
• Bilinguals show knowledge of their two L1s’ phonotactics, similar to monolinguals.
-
• Bilinguals differ from monolinguals in phonotactic processing.
1. Introduction
Adults have detailed knowledge of the phonotactic regularities of their language(s), that is, the legal and probabilistic positioning and sequencing of speech sounds within the words of their language(s). This phonotactic knowledge is mainly assessed via well-formedness judgment tasks, where individuals are asked to rate the wordlikeness of nonwords varying in phonotactic probabilities and are thus probed to think about the properties of their language metalinguistically. Such studies generally show that monolinguals provide ratings that correlate with the phonotactic regularities in their languages: the more frequent the regularities in the nonwords, the better their wordlikeness (e.g., Bailey & Hahn, Reference Bailey and Hahn2001; Frisch et al., Reference Frisch, Large, Zawaydeh, Pisoni, Bybee and Hopper2001; Hay et al., Reference Hay, Pierrehumbert, Beckman, Local, Ogden, Temple, Local, Ogden and Temple2004; Richtsmeier, Reference Richtsmeier2011; Vitevitch et al., Reference Vitevitch, Luce, Charles-Luce and Kemmerer1996). Besides this, phonotactic regularities have been shown to affect adult speech perception, facilitating speech processing. Phonotactic processing is studied with implicit tasks, designed to investigate the effect of phonotactics on online processing efficiency. Such studies have provided evidence that the frequency of segment sequences affects processing speed (Vitevitch & Luce, Reference Vitevitch and Luce1998, Reference Vitevitch and Luce2005). It also affects lexical access, as shown in lexical decision tasks in which listeners have to decide whether presented stimuli are real word in their language or not: monolingual speakers are quicker and more accurate to decide that stimuli are nonwords if they contain non-frequent phonotactic regularities (Vitevitch & Luce, Reference Vitevitch and Luce1999). Phonotactics also affects listeners’ speech segmentation (Boll-Avetisyan & Kager, Reference Boll-Avetisyan and Kager2016; Dumay et al., Reference Dumay, Frauenfelder and Content2002; McQueen, Reference McQueen1998; Mersad & Nazzi, Reference Mersad and Nazzi2011; van der Lugt, Reference Van der Lugt2001), syllable processing (Smith & Pitt, Reference Smith and Pitt1999), and phoneme categorization (Massaro & Cohen, Reference Massaro and Cohen1983; Pitt & McQueen, Reference Pitt and McQueen1998). Importantly, studies testing listeners with unfamiliar languages have shown that phonotactic effects in speech processing, shown by native speakers, do not appear with naïve listeners (e.g., Otake et al., Reference Otake, Yoneyama, Cutler and Van Der Lugt1996). In fact, phonotactic effects can result in “perceptual repairs” when processing non-native speech (e.g., Dupoux et al., Reference Dupoux, Parlato, Frota, Hirose and Peperkamp2011; Hallé et al., Reference Hallé, Dominguez, Cuetos and Segui2008). For example, in languages in which it is phonotactically illegal, speakers perceptually repair the onset /tl/ into their legal onset /kl/ (e.g., French: Hallé et al., Reference Hallé, Segui, Frauenfelder and Meunier1998; American-English: Hallé & Best, Reference Hallé and Best2007). Thus, these effects are not primarily signal-driven; rather, they appear language-specific and experience-driven.
The fact that phonotactics are acquired through experience raises the question of how bilinguals develop language-specific phonotactic knowledge and use it during processing. Indeed, bilinguals could be facing two challenges. First, they have to represent the phonotactic regularities of their two languages, even though some regularities might conflict between them (e.g., if a cluster is legal in one but illegal in other language). Second, they need to apply their knowledge dynamically as a function of language context, raising the question of how efficiently they activate/deactivate language-specific phonotactics during processing. These challenges might have particular impact on the representation and use of phonotactics in simultaneous bilinguals, who regularly use two languages they acquired from early infancy. In this study, we therefore aimed to investigate whether bilinguals know the phonotactics of both their languages (Experiment 1), and if so, whether they make equal use of such properties during speech processing (Experiment 2). To do so, we tested French–German simultaneous bilingual adults, and compared their behavior to monolingually-raised speakers of French and German.
Studies focussing on bilinguals’ phonotactics have mainly investigated sequential bilinguals who have acquired one language from birth and a second one later in lifeFootnote 1 (during childhood: Anisfeld et al., Reference Anisfeld, Anisfeld and Semogas1969; Altenberg & Cairns, Reference Altenberg and Cairns1983; Carlson et al., Reference Carlson, Goldrick, Blasingame and Fink2016; Frisch & Brea-Spahn, Reference Frisch and Brea-Spahn2010; Freeman et al., Reference Freeman, Blumenfeld and Marian2016, Reference Freeman, Blumenfeld, Carlson and Marian2022; Sebastián-Gallés & Bosch, Reference Sebastián-Gallés and Bosch2002; during adulthood: Boll-Avetisyan, Reference Boll-Avetisyan, Danis, Mesh and Sung2011; Lentz & Kager, Reference Lentz and Kager2015; Trapman & Kager, Reference Trapman and Kager2009; Weber & Cutler, Reference Weber and Cutler2006). Studies investigating phonotactic knowledge suggest that bilingual adults possess knowledge of both their phonotactic systems and that they are able to use it appropriately based on the identity of the language processed (Altenberg & Cairns, Reference Altenberg and Cairns1983; Frisch & Brea-Spahn, Reference Frisch and Brea-Spahn2010; but see Anisfeld et al., Reference Anisfeld, Anisfeld and Semogas1969). For example, Frisch and Brea-Spahn (Reference Frisch and Brea-Spahn2010) found that well-formedness judgments of spoken nonwords differing in phonotactic probability by early sequential Spanish-English bilingual adults did not differ significantly from those by monolingual speakers of English if nonwords were rated in English. Moreover, the bilinguals had nearly identical performance in wordlikeness ratings for both English and Spanish. This suggests that sequential bilinguals acquire both phonotactic systems at monolingual levels, and that such knowledge operates within each language independently during the judgment task.
Studies focusing on phonotactic processing suggest a more complex picture. On one hand, phonotactic effects are stronger when bilinguals process their dominant language, whether defined by greater language exposure in infancy/childhood, the language acquired first and/or language proficiency. For example, in a vowel detection gating task and an AX discrimination task, early sequential Spanish-English bilingual adults, like Spanish monolinguals, tend to repair word-initial sC clusters, which are unattested in Spanish, by adding a prosthetic /e/, but to a smaller extent and only when Spanish-dominant. Conversely, if they are English-dominant, they behave like English monolinguals and do not repair these clusters (Carlson et al., Reference Carlson, Goldrick, Blasingame and Fink2016). Moreover, Spanish-dominant early sequential Spanish-Catalan bilinguals show smaller sensitivity to Catalan phonotactics than Catalan-dominant bilinguals (Sebastián-Gallés & Bosch, Reference Sebastián-Gallés and Bosch2002). Such an effect of proficiency was further documented in late learners of Dutch, as Dutch proficiency related to the size of the phonotactic effect in a Dutch lexical decision task (Lentz & Kager, Reference Lentz and Kager2015; Trapman & Kager, Reference Trapman and Kager2009) and in a short-term memory recall task (Boll-Avetisyan, Reference Boll-Avetisyan, Danis, Mesh and Sung2011). On the other hand, sequential bilinguals show some interference of first language (L1) phonotactics during L2 processing, possibly because the L1 phonotactic system is not fully deactivated (Freeman et al., Reference Freeman, Blumenfeld and Marian2016, Reference Freeman, Blumenfeld, Carlson and Marian2022; Lentz & Kager, Reference Lentz and Kager2015; Weber & Cutler, Reference Weber and Cutler2006). For example, Weber and Cutler (Reference Weber and Cutler2006) showed that late sequential bilinguals, although they have acquired their L2 phonotactics, are not able to fully deactivate their L1 phonotactics. This interference effect of the L1 during L2 processing might also be related to the dominance effect described earlier: not only is phonotactic processing stronger in the dominant language, but perhaps it is harder for bilinguals to activate only L2 phonotactics during processing. An alternative explanation for these interference effects is that phonotactic processing is language-nonselective, which means that the bilinguals’ two phonotactic systems are always activated in parallel during processing. The latter explanation aligns with what is known about bilinguals’ lexical processing, which is also largely language-nonselective (e.g., De Groot & Nas, Reference De Groot and Nas1991; Marian & Spivey, Reference Marian and Spivey2003; Van Assche et al., Reference Van Assche, Brysbaert, Duyck, Heredia and Cieślicka2020).
Only one study tested both phonotactic knowledge and processing in sequential bilinguals. Altenberg and Cairns (Reference Altenberg and Cairns1983) investigated whether the phonotactic systems of German–English bilinguals, most of whom learned both languages before age 12 but were overall more proficient in English, interact during language processing. To test this, they used two tasks: a well-formedness judgment task and a lexical decision task, and presented participants with written nonwords that were phonotactically legal in one language while illegal in the other. Bilinguals were tested either in an English or in a German mode (i.e., the language to be considered for the task was either English or German, as were the instructions), and their performance in the English mode was compared with that of English monolinguals. In the judgment task, bilinguals showed language-specific adaptation: when tested in English, there was no evidence that their responses differed from those of English monolinguals. The authors interpreted this null result as suggesting that bilinguals’ phonotactic knowledge was correctly applied to the specific language context, with no cross-language interference. However, the lexical decision task revealed cross-language influence in phonotactic processing: bilinguals tested in English were slower than monolinguals to reject nonwords that were phonotactically legal in German but illegal in English. These results suggest that bilinguals’ phonotactic systems are not fully independent but rather remain jointly activated and can interact during processing: even the non-dominant language interfered with processing in the dominant language.
In summary, the above studies highlight two important aspects of sequential bilinguals’ phonotactics. First, sequential bilinguals’ experience does not appear to significantly influence their phonotactic knowledge in their two languages, as assessed with a metalinguistic judgment task, suggesting that their phonotactic representations are detailed and flexible enough to be selectively accessed to the appropriate language context. Second, the two phonotactic processing systems of sequential bilinguals do not seem to be completely independent from one another, as revealed by tasks requiring online processing and showing the influence of one language-specific system on the other. In addition to co-activation of the two phonotactic systems, sequential bilinguals also show evidence of language-dominance effects: the more proficient they are with one language, the stronger the phonotactic effects in processing that language.
Would the findings on phonotactic knowledge and processing extend to simultaneous bilinguals who were exposed to their two native languages from early infancy? Or could it be that more experience with the two languages, especially during a developmental period critical for phonological acquisition, helps better navigate between the two phonotactic systems during processing and mitigates dominance effects? To our knowledge, only one study focusing on phonotactic processing investigated simultaneous bilingual adults, comparing them to early and late sequential bilinguals, as well as monolingual controls (Parlato-Oliveira et al., Reference Parlato-Oliveira, Christophe, Hirose and Dupoux2010). This study investigated the identity of the epenthetic vowel repairing illegal consonant clusters in different groups of Japanese/Brazilian-Portuguese bilinguals (i.e., /u/ is the typical epenthetic vowel used by Japanese listeners, /i/ the typical vowel used by Brazilian listeners). It did not find clear differences between simultaneous and sequential bilinguals, who all patterned with the Brazilian monolinguals. The authors interpret these findings as resulting from bilinguals’ current linguistic proficiency and exposure, rather than language experience in infancy, given that all bilingual groups were Brazilian-dominant and living in Brazil. These findings suggest that current language exposure and use, rather than language input in childhood, influence phonotactic processing. Nevertheless, more studies are needed to better understand whether simultaneous bilinguals know the phonotactics of their two native languages and show interference of one language phonotactics when processing the other language. These aspects will be investigated in the current study.
1.1. The current study
In the present study, we used a similar design as Altenberg and Cairns (Reference Altenberg and Cairns1983) to investigate whether the dichotomy found in sequential bilinguals applies to simultaneous bilinguals, that is, phonotactic knowledge of their two languages being similar to monolinguals’ knowledge (Experiment 1), and phonotactic processing showing interference between their two languages (Experiment 2). Accordingly, bilinguals’ phonotactic knowledge was assessed through well-formedness judgments and phonotactic processing through lexical decisions, allowing to distinguish across tasks between phonotactic knowledge and online processing effects. The study’s materials were also inspired by their study, such that we manipulated the phonotactic properties of our stimuli in four ways: nonwords beginning with clusters that were (1) frequent or (2) non-frequent in both French and German, (3) frequent in French but non-frequent in German, or (4) frequent in German but non-frequent in French. This allowed us to assess whether phonotactic well-formedness in one language influences bilinguals’ behavior when perceiving the other language.
Besides our focus on simultaneous rather than sequential bilinguals, our study differed from Altenberg & Cairns’ study as follows. First, we presented auditory – as opposed to written – stimuli, given that a study found differences in phonotactic effects between the two modalities (Bailey & Hahn, Reference Bailey and Hahn2001). Second, we used stimuli pronounced by both a native speaker of French and a native speaker of German, to test whether pronunciation modulates the degree of cross-linguistic phonotactic activation in speech processing, given studies showing that language-specific pronunciations modulate bilinguals’ degree of parallel activation of their lexicons (Schulpen, Dijkstra, Schriefers & Hasper, Reference Schulpen, Dijkstra, Schriefers and Hasper2003; Lagrou, Hartsuiker, & Duyck, Reference Lagrou, Hartsuiker and Duyck2013; but see Lagrou, Hartsuiker, & Duyck, Reference Lagrou, Hartsuiker and Duyck2011). Third, we tested two monolingual-raised control groups, to validate our experiments and to directly test how bilinguals’ phonotactic effects compared to monolingual-raised speakers of their two languages. We initially expected that simultaneous bilinguals would show an effect of interference between their two languages in both drawing on knowledge (affecting their well-formedness judgments) and in using phonotactics for processing (affecting their lexical decisions). We are aware that for phonotactic knowledge, this hypothesis, which we included in our pre-registration, goes against the findings on sequential bilinguals reviewed above (Altenberg & Cairns, Reference Altenberg and Cairns1983; Frisch & Brea-Spahn, Reference Frisch and Brea-Spahn2010), and might be difficult to test with a metalinguistic judgment task capturing conscious knowledge rather than automatic processing, which limits its interpretative value regarding co-activation-based interference. However, we reasoned that it would be important to assess whether perhaps, with changes in design and methodology (the inclusion of simultaneous bilinguals and two monolingually-raised control groups, spoken stimuli and two pronunciations), we would find evidence for subtle interference between the bilinguals’ two L1s. Nevertheless, given the clearer evidence for phonotactic interference in processing than in drawing on knowledge found in sequential bilinguals, we expected that if task-specific effects arose, they would go in the same direction (i.e., interference would emerge in the lexical decision task). In terms of pronunciation cues, we hypothesized that if pronunciation cues play a role in the degree of cross-linguistic phonotactic activation, then the interference effect might be more likely to occur when language-specific pronunciations would not match the language being processed. These hypotheses were pre-registered on osf (well-formedness judgment task: https://osf.io/yda94; lexical decision task: https://osf.io/aep28) and included power analyses to justify our sample size for the bilingual group of participants, as well as the statistical analysis plans.
2. Experiment 1: Well-formedness judgment task
2.1. Methods
2.1.1. Participants
Thirty-one simultaneous French–German bilinguals (mean age = 26.39 years, range = 18–40 years, 21 females, 10 males), 30 French monolingually-raised (mean age = 29.53 years, range = 20–40 years, 25 females, five males), and 35 German monolingually-raised participants (mean age = 22.82 years, range = 18–34 years, 28 females, seven males) were tested and included in the analysis. None had hearing, language, or cognitive impairments, and all provided correct answers during the attention checks (see procedure below). Bilinguals had to be exposed to both French and German regularly before 2 years of age. Monolingually-raised participants were defined as individuals exclusively exposed to one language (French or German) from birth to at least 6 years of age. We did not collect information about languages learned later in life in the monolingually-raised groups, although formal foreign-language instruction is widespread in French and German schools. Thus, the term “monolingually-raised” refers to early monolingual language experience rather than the absence of later foreign-language knowledge. One additional bilingual was tested but excluded from the analyses due to experiment error. Eleven additional monolingually-raised participants were excluded from the analyses because they provided incorrect answers during the attention checks. A descriptive table characterizing the bilingual sample in language experience and proficiency can be found in Supplementary Appendix S2 (Table S3).
2.1.2. Materials
We created 64 nonwords divided into four sets differing in phonotactic properties. Two sets had convergent phonotactic properties across French and German, and two had divergent phonotactic properties between the two languages. Each nonword had a #CCVCVC# structure and had no direct lexical neighbor in French or German. Phonotactic differences between sets depended on the word-initial consonant cluster: in the FreqBoth set, the four word-initial clusters (/tr/, /dr/, /fl/, and /bl/) were frequent in both French and German. In the FreqNative/InfreqNative sets, the word-initial clusters were frequent in one language, but non-frequentFootnote 2 in the other. Hence, a stimulus in the FreqNative set for French was also in the InfreqNative set for German and vice versa, such that nonword status depended on the target language (see procedure below for more details). The four clusters phonotactically frequent in French while non-frequent in German were /st/, /sp/, /dj/, and /vr/. Conversely, the clusters that are frequent in German but non-frequent in French were /kn/, /kv/, /ʃp/, and /ʃt/. Finally, in the InfreqBoth set, word-initial clusters were non-frequent in both French and German: /ʃk/, /sr/, /sb/, and /fn/. Thus, four different word-initial clusters were used for each phonotactic set (16 in total). Each cluster was combined with four out of eight possible tails (i.e., /agik/, /opik/, /iton/, /orɛt/, /eʃɛt/, /usap/, /atyl/, /abɛt/; see Supplementary Appendix S1 for details about the stimuli) to create our final list of stimuli. Each tail was distributed evenly across the four phonotactic sets, ensuring that any difference between sets could only be explained by differences in phonotactic frequency of the word-initial clusters.
One female native speaker of French and one of German recorded all stimuli in the same soundproof booth using the same equipment. We normalized the stimuli’s mean intensity at 70 dB using PRAAT (Boersma, Reference Boersma2001). The average duration and pitch of nonwords for each phonotactic set and speaker are reported in Table 1. The French stimuli did not have lexical stress, which does not exist in French, whereas the German stimuli followed a strong-weak stress pattern, typical for German.
Average duration and pitch of the nonwords, divided by pronunciation and phonotactic group, as well as the real words used in Experiment 2
Table 1. Long description
The table is organized into five columns. The first column lists the phonotactic group or word type, while the remaining four columns are grouped under two main headers: French pronunciation and German pronunciation. Each pronunciation header contains sub-columns for Duration in seconds and Pitch in H z.
* Row 1, FreqBoth: French pronunciation 0.629 s, 241 H z; German pronunciation 0.644 s, 253 H z.
* Row 2, F R FreqNative and D E InfreqNative: French pronunciation 0.666 s, 238 H z; German pronunciation 0.750 s, 266 H z.
* Row 3, D E FreqNative and F R InfreqNative: French pronunciation 0.658 s, 224 H z; German pronunciation 0.740 s, 278 H z.
* Row 4, InfreqBoth: French pronunciation 0.680 s, 239 H z; German pronunciation 0.762 s, 273 H z.
* Row 5, German words: French pronunciation 0.599 s, 214 H z; German pronunciation 0.700 s, 226 H z.
* Row 6, French words: French pronunciation 0.641 s, 230 H z; German pronunciation 0.778 s, 231 H z.
2.1.3. Procedure, apparatus, and design
The experiment was conducted online using Labvanced (Finger et al., Reference Finger, Goeke, Diekamp, Standvoß and König2017).
Monolingual versions: Monolingually-raised participants were instructed that on each trial, they would hear sounds and were asked to rate how acceptable these sounds would be as a new word in French/German, using a scale from 1 (“Not acceptable at all”) to 7 (“Highly acceptable”). The experiment started with eight training trials and was followed by 128 test trials (i.e., the 64 nonwords in the two different pronunciations). A trial ended once the participant responded, or after 6 seconds. Because the task was web-based, four attention checks required participants to type the stimulus from the preceding trial. In these checks, responses were classified as incorrect when they indicated clear inattention (e.g., omitted responses or responses unrelated to the target item). Minor phonemic deviations (e.g., vowel substitutions or consonant changes within clusters) were not treated as errors, as such responses were taken to reflect successful perception of the stimulus despite small transcription differences. Pronunciation was blocked and presented in a fixed order, such that in the first half of the task (i.e., the first block) we presented the stimuli pronounced by the native speaker of the target language (target pronunciation) and in the second half of the task (i.e., the second block) we presented the same stimuli pronounced by the other speaker (non-target pronunciation). Trial order was pseudorandomized to avoid presenting more than two consecutive nonwords with the same word-initial cluster or more than four nonwords with the same tail in a sequence.
Bilingual version: Bilinguals completed this task in their two L1s. At the beginning, the procedure was exactly the same as for monolingually-raised participants: they completed the training and experimental trials and were asked to judge the acceptability of nonwords as new words of a given language, getting the instructions in that language. Once the participants had completed the 128 trials in that language, a written message stating that half of the task was completed was presented in both French and German. At this point, participants could take a break before resuming the experiment by pressing a key. Once resumed, the instructions appeared in the other language and prompted participants to judge the acceptability of nonwords as new words of that second language. We counterbalanced the order of target language across participants, such that half started with target language French while the other half started with target language German.
At the end of the experiment for the monolingually-raised participants, and after each language block for the bilinguals, participants completed the vocabulary questionnaire Lextale (German version: Lemhöfer & Broersma, Reference Lemhöfer and Broersma2012; French version: Brysbaert, Reference Brysbaert2013) assessing their proficiency in the tested language. Lextale consists of a list of written items (i.e., words and pseudowords) for which participants indicated whether they were existing words in the language. Finally, bilingual but not monolingually-raised participants were interviewed by phone on their language background (using a modified version of the LEAP-Q: Marian et al., Reference Marian, Blumenfeld and Kaushanskaya2007, mixed with the questionnaire used in Dupoux et al., Reference Dupoux, Peperkamp and Sebastián-Gallés2010, as in Boll-Avetisyan et al., Reference Boll-Avetisyan, Bhatara, Unger, Nazzi and Höhle2020) a few days after the experiment.
2.1.4. Predictions and data analysis
We predicted higher ratings for FreqBoth than InfreqBoth, and FreqNat than InfreqNat, such that the more frequent the phonotactic properties of the items, the higher their wordlikeness. Crucially, because we predicted phonotactic knowledge for both the monolingually-raised participants in their respective native languages, and the bilinguals in each of their languages, we predicted that both monolingually-raised participants and bilinguals would rate higher the frequent phonotactic conditions compared to the non-frequent conditions. Even though well-formedness judgments capture explicit, metalinguistic knowledge rather than automatic online processing (and are thus not a direct test of language co-activation), we nevertheless predicted that bilingual and monolingually-raised participants could differ in the size of phonotactic effects across conditions (notably between FreqNat and InfreqNat, where the phonotactic properties of their two languages diverged), and that language-specific pronunciation might differently affect monolingually-raised and bilingual listeners’ phonotactic well-formedness intuitions, attesting subtle differences in phonotactic knowledge.
To test those predictions, we analyzed participants’ rating in each trial with a (logit) ordinal mixed-effects model using the ordinal package (Christensen, Reference Christensen2019) in R (R Core Team, 2021; see Liddell & Kruschke, Reference Liddell and Kruschke2018 and the supplementary materials in Oh et al. Reference Oh, Todd, Beckner, Hay, King and Needle2020, for more details about this kind of analysis). Thus, the dependent variable was the participants’ rating on a Likert scale (1–7). The fixed part of the model included the main effects of three factors and their interactions: phonotactics (four levels, coded with sliding contrasts that compared InfreqBoth with FreqBoth, FreqBoth with FreqNative, and FreqNative with InfreqNative), pronunciation (two levels, sum-contrast coded, with an item being pronounced by the native versus non-native speaker of the target language, the target language being the one to which participants had to relate their wordlikeness ratings), and language status (two levels, monolingual versus bilingual, sum-contrast coded). The random structure was kept as maximal as possible (Barr et al., Reference Barr, Levy, Scheepers and Tily2013) and included by-participants random slopes for phonotactics in interaction with pronunciation, as well as by-items random slopes for pronunciation. The resulting formula was as follows:
Post-hoc comparisons of estimated marginal means (EMMs) were computed with the emmeans package (Lenth, Reference Lenth2024) and visualizations with the ggplot package (Wickham, Reference Wickham2016) in R.
Deviation from the pre-registration: When pre-registering the lexical decision task (Experiment 2), we realized that the pre-registered statistical model for Experiment 1 was not ideal. Our pre-registered model tested whether phonotactic properties, language-specific pronunciation, language status and their interaction affected participants’ ratings for each target language separately (French or German) with the following R-formula:
We provide the results of this model in Supplementary Appendix S3 (Table S7). However, considering our central hypotheses, this model was unnecessarily complex and difficult to interpret, as we were interested in the effect across the two languages. Thus, the present analyses follow the model pre-registered for Experiment 2. Furthermore, because our phonotactic factor has four levels, we initially planned to test the differences between FreqBoth-FreqNat, FreqNat-InfreqNat, and InfreqNat-InfreqBoth, expecting bilingualism effects only between FreqNat-InfreqNat. We later realized that testing differences between both FreqBoth-InfreqBoth and FreqNat-InfreqNat provided a more informative and concise analysis allowing more robust interpretations. Therefore, we used these contrasts here (and in Experiment 2, for which they were pre-registered).
2.2. Results
The results of the statistical model are presented in Table 2; a plot summary of the raw data is presented in Figure 1. These reveal that there were statistically significant main effects of the FreqBoth-InfreqBoth (p < .001) and InfreqNat-FreqNat (p < .001) contrasts: in both contrasts, participants rated the nonwords containing frequent clusters as more wordlike than nonwords containing non-frequent ones. There was also a main effect of pronunciation (p < .001), with participants rating the nonwords as more wordlike when pronounced by the native speaker of the target language. There was an interaction of phonotactics and pronunciation in both the FreqBoth-InfreqBoth (p < .001) and InfreqNat-FreqNat (p < .05) contrasts, indicating that the phonotactic effect was modulated by the native/non-native status of the speaker pronouncing the stimuli, with a larger effect if the nonwords were pronounced by the native speaker of the target language compared to when they sounded non-native. Finally, there was a significant interaction between the FreqBoth-InfreqBoth contrast and language status (p < .001), indicating that the phonotactic effect was smaller in bilinguals than monolingually-raised participants. Follow-up post-hoc tests per language group confirmed that both groups reliably distinguished the phonotactic conditions (the full output of orthogonal pairwise comparisons can be found in Supplementary Table S4). For monolingually-raised participants, FreqBoth was rated higher than InfreqBoth (log-odds = 1.45, OR = 4.26, p < .001) and FreqNat was rated higher than InfreqNat (log-odds = 1.12, OR = 3.05, p < .001). The same pattern was observed in the bilingual group, although with smaller effect sizes: FreqBoth was rated higher than InfreqBoth (log-odds = 0.78, OR = 2.19, p < .001), and FreqNat was rated higher than InfreqNat (log-odds = 0.87, OR = 2.38, p < .001). Thus, the effects of phonotactics were statistically significant in both monolingually-raised and bilingual participants, but the magnitude of the phonotactic effect also differed across groups, establishing subtle differences in phonotactic knowledge. No other main effects or interactions reached significance.
Ordinal mixed-effects model of well-formedness ratings, taking into account the phonotactic conditions, language-specific pronunciation, and language status
Table 2. Long description
The table is organized into five columns: Parameter, Estimate, Std. Error, z, and P. It is divided into two main sections: Effects and Thresholds.
Effects section:
* FreqBoth-Infreqboth: Estimate 1.115, Std. Error 0.145, z 7.692, P less than .001.
* FreqNat-Freqboth: Estimate minus 0.103, Std. Error 0.117, z minus 0.874, P .382.
* InfreqNat-FreqNat: Estimate minus 0.991, Std. Error 0.084, z minus 11.839, P less than .001.
* Pronunciation: Estimate 1.873, Std. Error 0.201, z 9.335, P less than .001.
* LangStatus: Estimate 0.266, Std. Error 0.274, z 0.972, P .331.
* FreqBoth-Infreqboth times pronunciation: Estimate 0.576, Std. Error 0.160, z 3.597, P less than .001.
* FreqNat-Freqboth times pronunciation: Estimate minus 0.143, Std. Error 0.139, z minus 1.025, P .305.
* InfreqNat-FreqNat times pronunciation: Estimate minus 0.212, Std. Error 0.099, z minus 2.143, P .032.
* FreqBoth-Infreqboth times LangStatus: Estimate minus 0.667, Std. Error 0.167, z minus 3.999, P less than .001.
* FreqNat-Freqboth times LangStatus: Estimate 0.201, Std. Error 0.114, z 1.760, P .078.
* InfreqNat-FreqNat times LangStatus: Estimate 0.248, Std. Error 0.164, z 1.517, P .129.
* pronunciation times LangStatus: Estimate minus 0.023, Std. Error 0.391, z minus 0.058, P .954.
* FreqBoth-Infreqboth times pronunciation times LangStatus: Estimate minus 0.099, Std. Error 0.192, z minus 0.517, P .605.
* FreqNat-Freqboth times pronunciation times LangStatus: Estimate minus 0.106, Std. Error 0.170, z minus 0.622, P .534.
* InfreqNat-FreqNat times pronunciation times LangStatus: Estimate 0.117, Std. Error 0.187, z 0.625, P .532.
Thresholds section (Parameter, Estimate, Std. Error):
* 1|2: minus 1.820, 0.145.
* 2|3: minus 0.183, 0.139.
* 3|4: 0.815, 0.139.
* 4|5: 1.468, 0.144.
* 5|6: 2.517, 0.145.
* 6|7: 4.030, 0.149.
Note: The two-level factors were coded with sum contrasts (i.e., pronunciation: non-target_pronun = −0.5/target_pronun = 0.5; LangStatus: Monolingual = −0.5/Bilingual = 0.5), while the factor phonotactic was coded with sliding contrasts (i.e., InfreqBoth. vs. FreqBoth, FreqBoth vs. FreqNative, and FreqNative vs. InfreqNative). *: p ≤ .05, **: p ≤ .01, ***: p ≤ .001.
Mean well-formedness ratings (and CIs) in relation to phonotactic condition (InfreqBoth/FreqBoth: clusters non-frequent/frequent in both French and German, respectively; FreqNat/InfreqNat: clusters frequent/non-frequent in the target language and non-frequent/frequent in the other language, respectively) split by language-specific pronunciation (target pronunciation, non-target pronunciation) and language status (monolingual, bilingual).
Figure 1. Long description
The graph consists of two side-by-side panels labeled Monolingual on the left and Bilingual on the right. The y-axis represents Mean Rating from 2.0 to 4.5. The x-axis for both panels lists four phonotactic conditions: InfreqBoth, FreqBoth, FreqNat, and InfreqNat. Two data series are plotted in each panel: target pronunciation (red circles with solid lines) and non-target pronunciation (teal triangles with dotted lines). Error bars represent C I s.
In the Monolingual panel:
- Target pronunciation starts at approximately 3.0 for InfreqBoth, peaks at 4.4 for FreqBoth, remains high at 4.2 for FreqNat, and drops to 3.2 for InfreqNat.
- Non-target pronunciation starts at 2.2 for InfreqBoth, rises to 2.8 for FreqBoth and FreqNat, and drops back to 2.2 for InfreqNat.
In the Bilingual panel:
- Target pronunciation starts at 3.5 for InfreqBoth, peaks at 4.3 for FreqBoth, stays high at 4.2 for FreqNat, and drops to 3.5 for InfreqNat.
- Non-target pronunciation starts at 2.5 for InfreqBoth, rises to 2.9 for FreqBoth and 3.0 for FreqNat, and drops to 2.4 for InfreqNat.
Across both panels, target pronunciation consistently receives higher ratings than non-target pronunciation. The bilingual group shows higher baseline ratings for the InfreqBoth and InfreqNat conditions compared to the monolingual group.
2.3. Discussion
Experiment 1 assessed simultaneous French–German bilinguals’ phonotactic knowledge with a well-formedness judgment task and compared it to that of monolingually-raised participants. Overall, our findings establish that, as predicted, all groups of participants were sensitive to the phonotactic properties of their native languages: the more frequent the phonotactic regularities in the nonwords, the better their wordlikeness. This is consistent with previous findings showing effects of language-specific phonotactic properties on well-formedness judgments by monolingual adults (e.g., Vitevitch et al., Reference Vitevitch, Luce, Charles-Luce and Kemmerer1996) and sequential bilinguals (Altenberg & Cairns, Reference Altenberg and Cairns1983; Frisch & Brea-Spahn, Reference Frisch and Brea-Spahn2010). Additionally, it shows that phonotactic differences restricted to word-initial clusters, as implemented here, are sufficient to elicit such effects. It also provides evidence that the task materials and procedure were valid and allowed the investigation of phonotactic knowledge in adults raised in monolingual or bilingual environments.
Turning to group comparisons, the results indicate that simultaneous bilinguals distinguished between the frequent and non-frequent phonotactic conditions in our task, as did monolingually-raised speakers. Only a small group difference emerged: a smaller effect in the bilingual compared to the monolingually-raised group in the InfreqBoth-FreqBoth contrast. While the source of this difference is unclear and might signal slight differences in (drawing on) phonotactic knowledge, it should be interpreted in light of the overall findings that bilinguals showed robust sensitivity to phonotactic patterns in both of their languages. These findings extend those found with sequential bilinguals, showing that their phonotactic knowledge is not significantly different from monolinguals (Altenberg & Cairns, Reference Altenberg and Cairns1983; Frisch & Brea-Spahn, Reference Frisch and Brea-Spahn2010). Thus, the hypothesis that co-activation would affect bilinguals’ phonotactic well-formedness intuitions was not supported by our data. We acknowledge that this was a difficult hypothesis to probe, given that the task was designed to capture conscious knowledge rather than automatic processing, limiting its interpretative value regarding co-activation-based interference.
Furthermore, we found that pronunciation cues modulated phonotactic well-formedness intuitions in both groups (smaller effects when items were pronounced by a speaker of the non-target language). However, we found no evidence that the language-specific pronunciation impacted bilinguals and monolingually-raised individuals differently. This suggests that pronunciation effects likely reflect general processing demands, with reduced effects under more demanding/unexpected pronunciations. We return to this effect in the General Discussion.
In sum, while Experiment 1 used a metalinguistic judgment task to establish that simultaneous bilinguals possess detailed phonotactic knowledge in both languages, Experiment 2 used a lexical decision task to investigate in the same bilinguals whether such knowledge is used during online phonotactic processing, and whether bilinguals show evidence of co-activation of their two phonotactic systems during processing.
3. Experiment 2: Lexical decision task
3.1. Methods
3.1.1. Participants
Twenty-three bilinguals (mean age = 25.7 years, range = 19-34 years, 13 females, 10 males) participated in the lexical decision task and were included in the analyses. They had participated in Experiment 1 at least 1 year before participating in Experiment 2. Additionally, 27 French monolingually-raised (mean age = 28.8 years, range = 18–40, 19 females, eight males) and 31 German monolingually-raised (mean age = 25.77, range = 18–40 years, 16 females, eight males, eight other/prefer not to say) participants who had not previously participated in Experiment 1 were tested and included in the analyses. Participants’ inclusion criteria were as in Experiment 1, except for the attention checks, which were not implemented in Experiment 2.
3.1.2. Materials
The nonword stimuli were a subset of the stimuli of Experiment 1Footnote 3 (32 out of 64 nonwords, see Supplementary Appendix S1 for more details). We also used 32 real German and 32 real French words, starting with some of the same word-initial clusters as the nonwords, as fillers. The words were pronounced by the same speakers who pronounced the nonwords, and audio files were processed as in Experiment 1. The average duration and pitch of the German and French words spoken by the French and the German native speaker can be found in Table 1.
3.1.3. Procedure, apparatus, and design
The experiment was conducted online using either Labvanced (Finger et al., Reference Finger, Goeke, Diekamp, Standvoß and König2017) for the monolingual French version or lab.js (Henninger et al., Reference Henninger, Shevchenko, Mertens, Kieslich and Hilbig2020) for the monolingual German and the bilingual versions. After the lexical decision task, monolingually-raised participants completed the language-appropriate LexTale vocabulary questionnaire (Brysbaert, Reference Brysbaert2013; Lemhöfer & Broersma, Reference Lemhöfer and Broersma2012). Bilinguals did not complete it again, having already done so for Experiment 1.
Monolingual versions: Participants were instructed that on each trial, they would hear a word and were asked to decide as quickly and accurately as possible whether what they heard was a real word or not by pressing a key. The task started with eight practice trials with feedback (“correct!”/“incorrect!” on screen). Following this, the test phase consisted of two blocks of 64 trials, half presenting real words (in German or French, depending on the target language) and half presenting nonwords. Within each block, trial order was pseudorandomized to avoid presenting more than two consecutive trials of the same lexical (i.e., word/nonword), phonotactic, or pronunciation condition.
Bilingual version: Bilinguals completed this task in their two L1s. At the beginning, the procedure was exactly the same as for monolingually-raised participants: after the practice phase, they were tested in one language, receiving instructions in that language. Once they had completed the 128 trials in that language, a written message stating that half of the task was completed was presented in both French and German. At this point, participants could take a break before resuming the experiment by pressing a key. Once resumed, they were tested and received instructions in the other language. We counterbalanced the order of target language across participants, such that half started with French while the other half started with German.
3.1.4. Predictions, data preprocessing, and analysis
We predicted that accuracy in rejecting a nonword as being a real word in the language would be higher for InfreqBoth than FreqBoth, and InfreqNat than FreqNat. We also predicted a significant interaction between language status and phonotactics, or between language status, phonotactics, and pronunciation, resulting from co-activation of bilinguals’ two phonotactic systems.
As pre-registered, before analyzing the data, we excluded trials during which participants responded too quickly (<100 ms; Vitevitch & Luce, Reference Vitevitch and Luce1999) or too slowly (>2500 ms; Lentz & Kager, Reference Lentz and Kager2015). Based on this criterion, 131 trials out of 6656 trials (1.97%) were excluded from the analysis. Furthermore, we excluded trials presenting real words, since those were only used as fillers in the task.
Our pre-registered analysis focused on the accuracy data of the nonwords with a (logit) generalized mixed-effects model. The independent factors and the model structure were the same as in Experiment 1. The random effects were grouped by participant and item. For this, we used R with the packages lme4 (Bates et al., Reference Bates, Mächler, Bolker and Walker2015) and lmerTest (Kuznetsova et al., Reference Kuznetsova, Brockhoff and Christensen2017) to obtain p-values. Because the model did not converge when including random slopes, we only included random intercepts for participants and items. The formula was as follows:
As suggested by a reviewer, in addition to the pre-registered accuracy model, we analyzed participants’ reaction times (RTs) in trials where they accurately rejected the nonwords (excluding 373–5.72% – inaccurate trials). Because the duration of the auditory stimuli was variable, RTs were corrected on a trial-by-trial basis by subtracting the duration of the spoken nonword from the raw response latency (e.g., McQueen, Reference McQueen1998). This correction provides estimates of response time relative to stimulus offset, thereby reducing variance attributable to differences in item durations. All subsequent analyses were conducted on these corrected RTs. Then, the RTs were screened for outliers using a model-based procedure that accounts for the hierarchical structure of the data. Specifically, we computed an intercept-only linear mixed-effects model including the full random-effects structure, with random intercepts and slopes for participants and items. This model captures subject/item-level variability in baseline response speed without accounting for experimental effects. Scaled residuals from this model were then computed, and observations with absolute residuals exceeding 2.5 standard deviations were excluded from further analyses (a total of 205 trials were excluded −3.33% – for a similar procedure, see Baayen & Milin, Reference Baayen and Milin2010). The proportion of excluded data was comparable across phonotactic conditions.
We used linear mixed-effects models to analyze RTs. The model structure was the same as for the accuracy model, except that random slopes for participants and items were included. The formula was as follows:
3.2. Results
3.2.1. Accuracy analyses
The results are presented in Table 3 and a plot summary of the raw data in Figure 2. They establish a statistically significant main effect of FreqBoth-InfreqBoth (p = .041) and InfreqNat-FreqNat (p < .001) indicating that participants were more accurate to decide that a stimulus was not a word in their language in the InfreqBoth than FreqBoth conditions, and the InfreqNat than FreqNat conditions. While the main effect of language status was not significant (p = .647), it significantly interacted with phonotactics (FreqBoth-InfreqBoth: p = .003; InfreqNat-FreqNat: p = .002), indicating that phonotactic properties modulated participants’ accuracy differently between the monolingually-raised and bilingual groups. Post-hoc comparisons (the full output of orthogonal pairwise comparisons can be found in Supplementary Table S5) revealed that monolingually-raised participants’ accuracy significantly differed between FreqBoth (EMM = 0.903, 95%CI [0.837; 0.945]) and InfreqBoth (EMM = 0.971, 95%CI [0.943; 0.985]; z = 3.117, p = .010), and between FreqNat (EMM = 0.920, 95%CI [0.872; 0.951]) and InfreqNat (EMM = 0.983, 95%CI [0.967; 0.991]; z = −5.872, p < .001). There were no significant differences between these conditions in the bilingual participants (FreqBoth: EMM = 0.958, 95%CI [0.913; 0.980], InfreqBoth: EMM = 0.968, 95%CI [0.932; 0.985], z = 0.656, p = 0.914; FreqNat: EMM = 0.947, 95%CI [0.901; 0.973], InfreqNat: EMM = 0.967, 95%CI [0.935; 0.984], z = −2.041, p = 0.173). Finally, there was a significant interaction between FreqBoth-FreqNative and pronunciation (p = .017); however, post-hoc comparisons failed to reveal significant accuracy differences between these two conditions in either pronunciation condition (stimuli pronounced by the speakers of the target [p = .791] or non-target language [p = .754]).
Generalized mixed-effects model of decision accuracy for nonwords, taking into account the phonotactic conditions, language-specific pronunciation, and language status
Table 3. Long description
The table contains 16 rows of statistical data organized into five columns: Parameter, Estimate, Std. Error, z, and p.
* (Intercept): Estimate 3.766, Std. Error 0.213, z 17.689, p < .001.
* FreqBoth minus InfreqBoth: Estimate -0.793, Std. Error 0.389, z -2.039, p .041.
* FreqNat minus FreqBoth: Estimate -0.016, Std. Error 0.334, z -0.047, p .962.
* InfreqNat minus FreqNat: Estimate 1.061, Std. Error 0.185, z 5.744, p < .001.
* Pronunciation: Estimate 0.139, Std. Error 0.126, z 1.104, p .270.
* LangStatus: Estimate 0.152, Std. Error 0.333, z 0.457, p .647.
* FreqBoth minus InfreqBoth times pronun: Estimate -0.099, Std. Error 0.344, z -0.288, p .773.
* FreqNat minus FreqBoth times pronun: Estimate -0.700, Std. Error 0.295, z -2.378, p .017.
* InfreqNat minus FreqNat times pronun: Estimate 0.054, Std. Error 0.368, z 0.148, p .883.
* FreqBoth minus InfreqBoth times LangStatus: Estimate 1.019, Std. Error 0.344, z 2.965, p .003.
* FreqNat minus FreqBoth times LangStatus: Estimate -0.464, Std. Error 0.295, z -1.574, p .116.
* InfreqNat minus FreqNat times LangStatus: Estimate -1.122, Std. Error 0.369, z -3.040, p .002.
* pronun times LangStatus: Estimate 0.346, Std. Error 0.252, z 1.373, p .170.
* FreqBoth minus InfreqBoth times pronun times LangStatus: Estimate 0.826, Std. Error 0.686, z 1.205, p .228.
* FreqNat minus FreqBoth times pronun times LangStatus: Estimate 0.048, Std. Error 0.589, z 0.081, p .935.
* InfreqNat minus FreqNat times pronun times LangStatus: Estimate -0.192, Std. Error 0.736, z -0.261, p .794.
Note: The two-level factors were coded with sum contrasts (i.e., pronunciation: non-target_pronun = −0.5/target_pronun = 0.5; LangStatus: Monolingual = −0.5/Bilingual = 0.5), while the factor phonotactic was coded with sliding contrasts (i.e., InfreqBoth vs. FreqBoth, FreqBoth vs. FreqNative, and FreqNative vs. InfreqNative). *: p ≤ .05, **: p ≤ .01, ***: p ≤ .001.
Summary of raw results from the lexical decision task, showing participants’ mean accuracy (and CIs) in rejecting the nonwords only in relation to phonotactic condition, split based on language status (monolingual, bilingual) and language-specific pronunciation (non-target pronunciation, target pronunciation).
Figure 2. Long description
The graph consists of two side-by-side panels with a shared Y-axis labeled Mean Accuracy ranging from 0.85 to 1.00. The X-axis is labeled phonotactics and contains four categories repeated in both panels: InfreqBoth, FreqBoth, FreqNative, and InfreqNative. A legend to the right indicates that red circles with solid lines represent target pronunciation and teal triangles with dotted lines represent non target pronunciation. Error bars indicate C I s.
* Monolingual Panel (Left): Accuracy for both groups drops sharply from InfreqBoth (approximately 0.96 to 0.98) to FreqBoth (approximately 0.90 to 0.91). For FreqNative, target pronunciation rises to 0.94 while non-target pronunciation stays low at 0.90. Both converge at a high point of 0.98 for InfreqNative.
* Bilingual Panel (Right): Trends are flatter compared to the monolingual group. Accuracy for both pronunciation types starts near 0.95 for InfreqBoth. For FreqBoth and FreqNative, target pronunciation accuracy dips slightly lower than non-target pronunciation, hovering between 0.92 and 0.93. Both groups converge again at 0.95 for InfreqNative.
Because the task was twice as long for bilinguals (two blocks of 128 trials) compared to monolingually-raised participants (one block of 128 trials), we explored whether any phonotactic effect in the lexical decision task for the bilinguals could have emerged as a function of block. Hence, we ran an additional generalized mixed-effects model with the bilingual data only, with three independent variables: phonotactics and pronunciation in interaction, nested within block. There was a significant main effect of block (β = −0.953, SE = 0.190, z = −5.018, p < .001), indicating that participants’ accuracy decreased between block1 and block2, and a significant main effect of FreqBoth-InfreqBoth in block2 (β = −0.789, SE = 0.377, z = −2.090, p = .037), indicating that accuracy was significantly higher in InfreqBoth (EMM = 0.958; 95%CI [0.907; 0.982]) compared to FreqBoth (EMM = 0.913; 95%CI [0.830; 0.959]; see Figure 3 and Table 4).
Summary of raw results from the lexical decision task, showing bilingual participants’ mean accuracy (and CIs) for the nonwords only, in relation to phonotactic condition, split based on Block (Block1, Block2).
Figure 3. Long description
The y-axis represents Mean Accuracy ranging from 0.85 to 1.00. The x-axis is labeled phonotactics and contains four categories: InfreqBoth, FreqBoth, FreqNative, and InfreqNative. A legend to the right identifies two series: Block 1 represented by red circles and a solid line, and Block 2 represented by teal triangles and a dashed line. Vertical error bars indicate C I s for each data point.
* At InfreqBoth: Block 1 accuracy is approximately 0.955 and Block 2 is approximately 0.945.
* At FreqBoth: Block 1 increases to approximately 0.97 while Block 2 drops significantly to approximately 0.90.
* At FreqNative: Block 1 decreases to approximately 0.95 and Block 2 remains stable at approximately 0.90.
* At InfreqNative: Block 1 increases to approximately 0.965 and Block 2 increases to approximately 0.93.
Overall, Block 1 maintains higher accuracy across all conditions compared to Block 2, with the largest performance gap occurring in the FreqBoth and FreqNative conditions.
Results of the exploratory generalized mixed-effects model assessing the effect of phonotactics and pronunciation within each block on simultaneous bilingual participants’ accuracy in the lexical decision task (nonwords only)
Table 4. Long description
The table contains 16 rows of statistical data organized into five columns: Parameter, Estimate, Std. Error, z, and p.
* (Intercept): Estimate 3.921, Std. Error 0.355, z 11.046, p less than .001.
* Block: Estimate minus 0.953, Std. Error 0.190, z minus 5.018, p less than .001.
* Block 1 FreqBoth minus InfreqBoth: Estimate 0.551, Std. Error 0.482, z 1.142, p .253.
* Block 2 FreqBoth minus InfreqBoth: Estimate minus 0.789, Std. Error 0.377, z minus 2.090, p .037.
* Block 1 FreqNat minus FreqBoth: Estimate minus 0.565, Std. Error 0.470, z minus 1.201, p .230.
* Block 2 FreqNat minus FreqBoth: Estimate minus 0.023, Std. Error 0.335, z minus 0.069, p .945.
* Block 1 InfreqNat minus FreqNat: Estimate 0.337, Std. Error 0.430, z 0.784, p .433.
* Block 2 InfreqNat minus FreqNat: Estimate 0.608, Std. Error 0.325, z 1.868, p .062.
* Block 1 pronunciation: Estimate 0.076, Std. Error 0.302, z 0.252, p .801.
* Block 2 pronunciation: Estimate 0.408, Std. Error 0.224, z 1.824, p .068.
* Block 1 FreqBoth minus InfreqBoth times pronun: Estimate minus 0.093, Std. Error 0.884, z minus 0.105, p .916.
* Block 2 FreqBoth minus InfreqBoth times pronun: Estimate 0.520, Std. Error 0.648, z 0.802, p .422.
* Block 1 FreqNat minus FreqBoth times pronun: Estimate minus 0.997, Std. Error 0.871, z minus 1.144, p .252.
* Block 2 FreqNat minus FreqBoth times pronun: Estimate minus 0.409, Std. Error 0.574, z minus 0.713, p .476.
* Block 1 InfreqNat minus FreqNat times pronun: Estimate 0.100, Std. Error 0.825, z 0.121, p .904.
* Block 2 InfreqNat minus FreqNat times pronun: Estimate minus 0.134, Std. Error 0.615, z minus 0.218, p .827.
3.2.2. RT analyses
The results are presented in Table 5 and a plot summary of the raw data in Figure 4. They establish a significant main effect of FreqBoth-InfreqBoth (p = .041) and InfreqNat-FreqNat (p < .001) indicating that participants were faster in accurately deciding that a stimulus was not a word in their language in the InfreqBoth than FreqBoth conditions, and the InfreqNat than FreqNat conditions. While the main effect of language status was not significant (p = .153), it significantly interacted with phonotactics (FreqBoth-InfreqBoth: p = .002; InfreqNat-FreqNat: p = .001), indicating that phonotactic properties modulated participants’ RTs differently between the monolingually-raised and bilingual groups. Post-hoc comparisons (the full output of orthogonal pairwise comparisons can be found in Supplementary Table S6) revealed that monolingually-raised participants’ RTs significantly differed between FreqBoth (EMM = 514 ms, 95%CI [452 ms; 575 ms]) and InfreqBoth (EMM = 435 ms, 95%CI [371 ms; 498 ms]; z = 2.964, p = .009), and between FreqNat (EMM = 515 ms, 95%CI [453 ms; 577 ms]) and InfreqNat (EMM = 418 ms, 95%CI [359 ms; 477 ms]; z = −7.872, p < .001). There were no significant differences between these conditions in the bilingual participants (FreqBoth: EMM = 418 ms, 95%CI [330 ms; 507 ms], InfreqBoth: EMM = 389 ms, 95%CI [297 ms; 481 ms], z = 1.012, p = 0.630; FreqNat: EMM = 404 ms, 95%CI [311 ms; 496 ms], InfreqNat: EMM = 379 ms, 95%CI [291 ms; 467 ms], z = −1.534, p = 0.302). Finally, there was a significant interaction between pronunciation and InfreqNat-FreqNat (p = .002), indicating that, overall, the difference in RTs between these two phonotactic levels was bigger in the non-target pronunciation (RT-diff = −84.08 ms, SE = 12.6 ms, z = −6.653, p < .001) compared to the target pronunciation (RT-diff = −37.57 ms, SE = 12.6 ms, z = −2.989, p = .008). No other effects reached significance.
Results of the mixed-effects model of corrected RTs in accurately rejecting nonwords in Experiment 2, taking into account the phonotactic conditions, language-specific pronunciation, and language status
Table 5. Long description
The table contains six columns: Parameter, Estimate, Std. Error, t, and p.
* (Intercept): Estimate 433.895, Std. Error 26.605, t 16.309, p less than .001.
* FreqBoth minus InfreqBoth: Estimate 54.375, Std. Error 25.704, t 2.115, p .041.
* FreqNat minus FreqBoth: Estimate minus 6.455, Std. Error 21.721, t minus 0.297, p .768.
* InfreqNat minus FreqNat: Estimate minus 60.826, Std. Error 10.189, t minus 5.970, p less than .001.
* Pronunciation: Estimate minus 1.578, Std. Error 14.831, t minus 0.106, p .916.
* LangStatus: Estimate minus 73.046, Std. Error 50.628, t minus 1.443, p .153.
* FreqBoth minus InfreqBoth times pronun: Estimate minus 1.131, Std. Error 41.961, t minus 0.027, p .979.
* FreqNat minus FreqBoth times pronun: Estimate 30.253, Std. Error 37.152, t 0.814, p .421.
* InfreqNat minus FreqNat times pronun: Estimate minus 46.517, Std. Error 14.836, t minus 3.135, p .002.
* FreqBoth minus InfreqBoth times LangStatus: Estimate minus 49.449, Std. Error 22.305, t minus 2.217, p .030.
* FreqNat minus FreqBoth times LangStatus: Estimate minus 15.978, Std. Error 16.661, t minus 0.959, p .339.
* InfreqNat minus FreqNat times LangStatus: Estimate 71.807, Std. Error 20.368, t 3.525, p less than .001.
* Pronun times LangStatus: Estimate minus 12.738, Std. Error 10.500, t minus 1.213, p .225.
* FreqBoth minus InfreqBoth times pronun times LangStatus: Estimate minus 19.412, Std. Error 29.767, t minus 0.652, p .514.
* FreqNat minus FreqBoth times pronun times LangStatus: Estimate 38.890, Std. Error 30.011, t 1.296, p .195.
* InfreqNat minus FreqNat times pronun times LangStatus: Estimate minus 13.335, Std. Error 29.644, t minus 0.450, p .653.
Summary of raw results from the lexical decision task, showing participants’ mean corrected reaction times (and CIs) in accurately rejecting the nonwords in relation to phonotactic condition, split based on language status (monolingual, bilingual).
Figure 4. Long description
The y-axis represents Mean Corrected R T in milliseconds, ranging from 350 to 550. The x-axis is labeled phonotactics and contains four categories: InfreqBoth, FreqBoth, FreqNative, and InfreqNative. A legend to the right identifies two groups: Monolingual represented by a solid red line with circular markers, and Bilingual represented by a dashed teal line with triangular markers. Both lines include vertical error bars representing confidence intervals.
Data trends from left to right:
* At InfreqBoth, Monolinguals are at approximately 425 milliseconds and Bilinguals are lower at approximately 390 milliseconds.
* At FreqBoth, Monolinguals show a sharp increase to approximately 490 milliseconds, while Bilinguals show a smaller increase to approximately 425 milliseconds.
* At FreqNative, Monolinguals peak at approximately 500 milliseconds, while Bilinguals decrease slightly to approximately 410 milliseconds.
* At InfreqNative, both groups decrease significantly. Monolinguals drop to approximately 415 milliseconds and Bilinguals drop to approximately 385 milliseconds.
Overall, the Monolingual group consistently maintains higher reaction times across all conditions compared to the Bilingual group, with the largest disparity occurring in the FreqBoth and FreqNative conditions.
Just as for the accuracy analyses, we explored whether any phonotactic effect in the lexical decision task for the bilinguals could have emerged as a function of block. Hence, we ran an additional linear mixed-effects model with the bilingual data only, with three independent variables: phonotactics and pronunciation in interaction nested within block. There was a significant main effect of block (β = −47.495, SE = 8.286, t = −5.732, p < .001), indicating that participants’ RTs decreased between block1 and block2, a significant main effect of FreqNat-InfreqNat in block2 (β = −36.828, SE = 16.650, t = −2.212, p = .027, Figure 5 and Supplementary Table S2), indicating that RTs were smaller in InfreqNat (EMM = 349 ms, 95%CI [277 ms; 421 ms]) compared to FreqNat (EMM = 386 ms, 95%CI [314 ms; 458 ms]), and a significant interaction between FreqNat-InfreqNat and pronunciation in block1, indicating that the phonotactic effect differed between pronunciation conditions in the first block. However, post-hoc analyses showed that the InfreqNat-FreqNat contrast was not significant in either the native (p = .708) or non-native (p = .156) condition when considered separately. This pattern indicates that the interaction arises from a difference between conditions rather than from a robust effect within either condition alone.
Summary of raw results from the lexical decision task, showing participants’ mean corrected reaction times (and CIs) in accurately rejecting the nonwords in relation to phonotactic condition, split based on block (Block1, Block2).
Figure 5. Long description
The Y-axis represents Mean Corrected R T in milliseconds, ranging from 300 to 550. The X-axis lists four phonotactic conditions: InfreqBoth, FreqBoth, FreqNative, and InfreqNative. Two data series are plotted: Block 1 (solid red line with circular markers) and Block 2 (dashed teal line with triangular markers). Both series include vertical error bars representing confidence intervals.
* At InfreqBoth, Block 1 is approximately 415 ms and Block 2 is approximately 385 ms.
* At FreqBoth, both series peak. Block 1 rises to approximately 450 ms and Block 2 rises to approximately 410 ms.
* At FreqNative, both series show a slight decrease. Block 1 is approximately 430 ms and Block 2 is approximately 395 ms.
* At InfreqNative, Block 1 remains relatively stable at approximately 425 ms, while Block 2 shows a sharp decline to its lowest point at approximately 355 ms.
Overall, Block 2 consistently shows faster reaction times than Block 1 across all conditions, with the largest gap occurring at the InfreqNative condition.
3.3. Discussion
Experiment 2 probed simultaneous French–German bilinguals’ phonotactic processing with a web-based lexical decision task, and compared their data to monolingually-raised participants. Overall, findings show that, as predicted, phonotactic properties modulated participants’ lexical decisions: accuracy in rejecting a nonword as a real word of the native language was significantly better for nonwords composed of non-frequent compared to frequent regularities, entailing that the more frequent the phonotactic regularities, the more often a nonword was confused with a real French/German word. However, these effects were mainly driven by the monolingually-raised groups. Indeed, monolingually-raised and bilingual participants significantly differed in how phonotactic properties modulated their decisions: while significant effects of phonotactics on accuracy and RTs were found in monolingually-raised participants (in both the InfreqBoth-FreqBoth and FreqNat-InfreqNat levels), the same effects were not found in bilinguals. Only block analyses in bilinguals revealed phonotactic effects in the second half of the experiment, though restricted to the FreqBoth-InfreqBoth levels for accuracy, where the phonotactic properties of their two languages converged, and the FreqNat-InfreqNat levels for RTs, where they diverged.
These findings extend previous language-specific phonotactic processing effects in monolinguals (e.g., Vitevitch & Luce, Reference Vitevitch and Luce1999) to French and German, providing evidence that the task materials and web-based procedure were valid. In light of weak phonotactic effects in the bilingual group, the difference between monolingually-raised and bilingual participants’ phonotactic processing suggests that bilinguals may rely differently on phonotactic properties during online processing than monolingually-raised individuals. One possible interpretation is that having to navigate between two phonotactic systems modulates how strongly language-specific phonotactic regularities are involved during lexical decisions. We return to these findings in the General Discussion.
4. General discussion
This study investigated phonotactic knowledge and processing in simultaneous French–German bilinguals and compared them to monolingually-raised speakers of these languages in two experiments: a metalinguistic well-formedness judgment task (Experiment 1) and a lexical decision task (Experiment 2). Because only one prior study assessed simultaneous bilinguals’ phonotactics (Parlato-Oliveira et al., Reference Parlato-Oliveira, Christophe, Hirose and Dupoux2010), our first aim was to investigate whether, and if so how, their phonotactic knowledge and use of it during online processing would differ from two groups of monolingually-raised participants. Second, we investigated whether bilinguals’ two phonotactic systems interact, potentially leading to differences from the monolingually-raised groups. To do so, we manipulated the phonotactic properties of our stimuli following Altenberg and Cairns (Reference Altenberg and Cairns1983), such that nonwords started with word-initial clusters that were either frequent or non-frequent in both French and German (convergent phonotactics), or frequent in French but non-frequent in German or vice versa (divergent phonotactics), allowing us to directly assess whether phonotactic well-formedness in one language could influence bilinguals’ behavior when processing the other. Furthermore, our stimuli were pronounced by both a native speaker of French and German, to explore the potential modulation of language interference by language-specific pronunciation factors.
Starting with the monolingually-raised individuals, our results in the two tasks demonstrated, as predicted, that phonotactic properties played a role in their performance. In the well-formedness judgment task, ratings were significantly higher for nonwords composed of frequent compared to non-frequent phonotactic regularities, such that more frequent phonotactic properties yielded higher wordlikeness ratings. In the lexical decision task, accuracy and RTs were significantly more accurate and faster for nonwords composed of non-frequent compared to frequent regularities. Results also showed that language-specific pronunciation modulated both monolingually-raised participants’ overall ratings (with lower ratings in the non-target pronunciation condition) and the phonotactic effect in the judgment task, but not the lexical decision task, with a smaller but significant phonotactic effect when the items were pronounced by a speaker of the non-target language. This finding could imply that pronunciation properties have a strong influence on participants’ wordlikeness judgments, and also that they interfere with the strength of phonotactic intuitions. However, methodological choices in Experiment 1 might have prompted such an effect of pronunciation. Indeed, it is possible that ratings were overall lower in the non-target pronunciation because of our task instructions: had we explicitly said to abstract away pronunciation cues, perhaps we would not have found such an effect. Furthermore, participants were presented first with pronunciations of the target language followed by pronunciations of the non-target language, in a fixed order, which could have exacerbated differences in judgments between the two conditions. This might then also explain why pronunciation did not affect lexical decisions. Beyond the main effect of pronunciation, the fact that it influenced the magnitude of the phonotactic effect might result from more demanding processing for atypical pronunciations, with phoneme categorization being more impacted, which could have cascading effects impacting phonotactic and lexical processing (Floccia et al., Reference Floccia, Butler, Goslin and Ellis2009; Williams & Escudero, Reference Williams and Escudero2014). However, it should be noted that the effect of pronunciation was also found with the bilinguals. Thus, this finding should be interpreted with caution and be replicated in other experiments, as it did not emerge in the lexical decision task, possibly because processing at the lexical level is more resilient to sub-lexical pronunciation properties. However, task-specific effects could also explain such differences across the two experiments, since sub-lexical properties are not the explicit focus of the lexical decision task, and items pronounced by both speakers were presented together in a random order.
Turning to the simultaneous bilinguals, our main group of interest, we found a clear demonstration of their knowledge of the phonotactic properties of their two native languages in the well-formedness judgment task, extending to simultaneous bilingualism previous findings on sequential bilingualism (e.g., Altenberg & Cairns, Reference Altenberg and Cairns1983; Frisch & Brea-Spahn, Reference Frisch and Brea-Spahn2010). However, findings from the lexical decision task suggest differences between bilingual and monolingually-raised individuals in their use of phonotactic knowledge during processing.
Their performance in the well-formedness judgment task suggests that indeed, simultaneous bilinguals apply phonotactic knowledge appropriately to the language being processed even when asked to shift between their languages within the same task. This suggests that bilinguals were operating in language-specific modes in this rating task, and that they possess detailed phonotactic knowledge in both of their languages. This is further supported by the fact that pronunciation interfered with phonotactic wordlikeness in the same way in the bilingual and monolingually-raised groups, even though the former group should have more experience with both pronunciations. Nevertheless, ratings in bilinguals were not modulated by phonotactic properties in the exact same way as in monolingually-raised participants, as revealed by one (and only one) difference between groups, namely a significantly smaller difference in ratings between the FreqBoth-InfreqBoth levels in the bilingual compared to the monolingually-raised group. However, this difference did not extend to the FreqNat-InfreqNat levels, suggesting that this effect is small and not robust. Moreover, the fact that such a difference emerged only in levels for which the phonotactic properties converged across French and German (i.e., regularities were either frequent or non-frequent in both languages) and not in levels where the two languages diverged phonotactically makes it unlikely to result from language-nonselective activation of bilinguals’ phonotactic representations. Rather, it is possible that it results from a group difference in acceptability thresholds, with simultaneous bilinguals being more willing to accept any sequence of sounds as new words in their languages, perhaps because they are more regularly exposed to language variations at various levels (see Bialystok, Reference Bialystok2001; for a review of metalinguistic flexibility in bilingual children).
Findings in the lexical decision task, which evaluates phonotactic processing rather than knowledge, suggest a completely different picture, showing clear differences between the bilingual and monolingually-raised participants. Whereas monolingually-raised participants’ accuracy and RTs significantly differed as a function of phonotactics (in both the InfreqBoth-FreqBoth and FreqNat-InfreqNat levels), no significant overall phonotactic effects were found in the bilinguals. However, an additional analysis comparing bilinguals’ performance between the two blocks of the experiment (each block corresponding to one of their languages) showed that bilinguals’ performance was modulated by phonotactics only in the second block, with a significant difference in accuracy between the levels FreqBoth-InfreqBoth and a significant difference in RTs between the levels FreqNat-InfreqNat. These effects might reflect both task-specific factors and interactions between bilinguals’ two phonotactic systems. In the first block, the lack of a significant phonotactic effect may relate to ceiling-level accuracy and relatively slow decisions, while in the second block, participants’ accuracy and RTs significantly decreased (e.g., for accuracy, from 96% to 92%; for RTs, from 430 to 386 ms). Because stimuli in both pronunciations occurred within a block, it might have been harder to efficiently process the stimuli’s phonotactics compared to Experiment 1, requiring bilinguals to make more effort to suppress the activation of words in the other language and prompting them to be particularly careful in their decisions in the first half of the experiment to cope with such unexpected variation. It is only once familiarized with the task that some effects emerged, which coincided with more errors and quicker responses. Nevertheless, the significant phonotactic effects found in the second block differed between accuracy and RT analyses. For accuracy, our results would fit with a language nonselective activation account, which might have modulated the phonotactic effect: the smaller, non-significant, difference between the FreqNat and InfreqNat levels could result from interference in phonotactic processing when the phonotactic properties diverged in the bilinguals’ two native languages, while the bigger, significant, difference between the FreqBoth and InfreqBoth levels could result from a strengthened phonotactic effect when the phonotactic properties converged in their two native languages. If our interpretation is correct, this effect would (weakly) replicate, and extend to simultaneous bilinguals, what Altenberg and Cairns (Reference Altenberg and Cairns1983) found with sequential German–English bilinguals using written stimuli: in their study, sequential bilinguals were slower than English monolinguals to reject nonwords as real words of English when these nonwords were composed of legal phonotactic regularities in German that were illegal in English, suggesting interference of German phonotactics when processing English. The fact that in our study such effects emerged only in lexical decisions but not well-formedness judgments is consistent with the idea that online processing tasks are more sensitive to cross-language interactions than metalinguistic tasks (Freeman et al., Reference Freeman, Blumenfeld, Carlson and Marian2022). However, this effect only emerged in the second half of Experiment 2 and was not confirmed in the RT analyses. In the RT analysis, the only significant effect emerging in block2 was specific to the FreqNat-InfreqNat levels, with quicker decisions in InfreqNat compared to FreqNat. Thus, the language nonselective interpretation of bilinguals’ phonotactic processing should be taken cautiously and replicated before strong conclusions are drawn about the mechanisms underlying such group differences. Alternatively, it is also possible that co-activation of their two phonotactic systems is reduced in simultaneous bilinguals, as opposed to sequential bilinguals, because of smaller asymmetries in proficiency/knowledge between their two languages (Sebastián-Gallés et al., Reference Sebastián-Gallés, Echeverría and Bosch2005; but see Luk & Bialystok, Reference Luk and Bialystok2013), or because of smaller L1 interference effects (Flege, Reference Flege, Schiller and Meyer2003; Iverson et al., Reference Iverson, Kuhl, Akahane-Yamadac, Dieschd, Tohkurae, Kettermannf and Siebert2003). If so, L1 interference during L2 processing might be weaker in bilingually-raised individuals. Studies directly comparing these populations will be important to clarify this issue.
In summary, Experiment 1 suggests that simultaneous bilinguals have knowledge of the phonotactics of both native languages comparable to that of monolingually-raised individuals and can use such knowledge offline in a language-specific way. Moreover, Experiment 2 suggests that bilinguals process phonotactics differently from monolingually-raised individuals, possibly due to cross-language interactions during online processing, although this requires further investigation.
Before concluding, we would like to describe some limitations of the current study. One concerns the language background of the monolingually-raised control groups. Participants were required to have had exposure to only one language during early childhood; however, we did not collect information about languages learned later in life. Because foreign-language instruction is widespread from primary school onward in the French and German educational systems, it is likely that many participants had some classroom-based knowledge of foreign language (most commonly English or Spanish). Importantly, such experience differs substantially from early simultaneous bilingual exposure and does not entail relatively balanced use of two phonotactic systems from infancy. Furthermore, if any participants had acquired knowledge of the other target language, this would likely have reduced differences between monolingually-raised and bilingual groups, making our comparison more conservative. Therefore, we acknowledge that systematically documenting later-acquired language experience in our control groups would have further informed group comparisons. A second limitation relates to the cross-task comparison design implemented to investigate bilinguals’ phonotactic knowledge (Experiment 1) and processing (Experiment 2) and how the two relate. The two tasks differed in their structure and dependent measures: the judgment task used a Likert scale, whereas the lexical decision task relied on accuracy/RTs. These differences in response measures might influence statistical sensitivity and thus limit the strength of conclusions stemming from direct comparisons across the two tasks. Nevertheless, we believe that by clearly defining the construct investigated with each task (knowledge vs. processing), we refrained from making direct comparisons. A third limitation concerns the pronunciation manipulation, as only one speaker per language was used to pronounce our items, raising the possibility that some effects result from idiosyncratic speaker characteristics. However, the pattern of results was symmetric across language groups suggesting that the findings are unlikely to be driven only by speaker-specific properties (e.g., in Experiment 1, French monolingually-raised participants showed reduced effects for German pronunciations and vice versa, although this is not explicitly apparent in our models because the two monolingually-raised groups were collapsed together). Finally, a fourth limitation relates to the web-based nature of the present study. A clear advantage of such a design is the facilitation of recruitment and testing of uncommon populations (in our case, simultaneous French–German bilinguals). A clear disadvantage is that the lack of experimental control over both participants (e.g., mental states, settings, and levels of external distraction during the experiment) and technical equipment often results in noisier data, which might have reduced statistical power to reliably detect some effects. While our power analyses indicated that sample sizes were sufficiently large to detect phonotactic effects in both tasks, the data might have been underpowered for RT analyses in Experiment 2, which is why only accuracy analyses were pre-registered. Replication under controlled laboratory settings would therefore be useful.
In conclusion, our study showed, with two tasks, that individuals raised in a bilingual French–German environment (i.e., simultaneous bilinguals) have knowledge of the phonotactic properties of their two languages at similar level as monolingually-raised individuals, and are able to activate such knowledge selectively. Nevertheless, our findings suggest that bilinguals’ phonotactic processing differs from that of monolingually-raised individuals in their two native languages, possibly because of co-activation of phonotactic representations in their two languages, resulting in interference when properties conflict across languages. Our study opens up the question of whether simultaneous bilinguals differ from sequential bilinguals in the degree of language-selective activation of phonotactics in their two languages. While most studies showed an important effect of L1 phonotactic knowledge during L2 processing in sequential bilinguals (but see Altenberg & Cairns, Reference Altenberg and Cairns1983 for the opposite direction), cross-language effects were less clear and robust in our study, which could imply greater abilities in navigating between their two languages’ phonotactics in simultaneous bilinguals.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S1366728926101667.
Data availability statement
The data and the R-script for data analysis are available on the following open repository in OSF: https://osf.io/kqu48/overview?view_only=8786b2b306e44ccfb066d96eb2a20943
Acknowledgements
We thank all the persons who participated in the study. A special thanks to Ekaterina Artemewa, Vladislav Trakal, and Yulia Pachenko for their precious help in conducting this study.
Author contribution
Conceptualization: L.P., T.N., N.B.-A.; Data curation: L.P.; Formal analysis: L.P., T.N., N.B.-A.; Funding acquisition: T.N., N.B.-A; Investigation: L.P.; Methodology: L.P., T.N., N.B.-A.; Project administration: L.P., T.N., N.B.-A.; Resources: T.N., N.B.-A.; Supervision: T. N., N.B.-A; Validation: T.N., N.B.-A.; Visualization: L.P.; Writing–original draft & revision: L.P.; Writing– revision & editing: T.N., N.B.-A.
Funding statement
This work was partially funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – Project ID 317633480 – SFB 1287, Project C07; and by a public grant overseen by the IdEx Université Paris Cité (ANR-18-IDEX-0001) as part of the Labex Empirical Foundations of Linguistics – EFL.
Competing interests
The authors declare none.
Ethical statement
The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008.
Open access
