Interference between non-native languages during trilingual language production

Journal of Memory and Language 128 (2023) 104386 Available online 9 November 2022 0749-596X/© 2022 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/). Interference between non-native languages during trilingual language production Angela de Bruin a , * , Liv J. Hoversten b , Clara D. Martin c , d a Department of Psychology, University of York, York, UK b Department of Psychology, University of California, Santa Cruz, USA c Basque Center on Cognition, Brain and Language (BCBL), Donostia-San Sebasti´ an, Spain d Ikerbasque, Basque Foundation for Science, Bilbao, Spain ARTICLE INFO Keywords: Trilingualism Language production Language interference Inhibition ABSTRACT Most research on multilingual language control has focused on a bilingual’s first (L1) and second (L2) languages. Studies on third language (L3) acquisition suggest that, despite the L1 being more proficient, L3 learners experience more L2 than L1 interference. However, little is known about how a trilingual’s L2 and L3 interact after initial stages of language learning. In the current study (Experiment 1: 30 Spanish-Basque-English trilinguals; Experiment 2: 50 English-French-Spanish trilinguals), participants completed a speeded naming task to assess cross-language intrusions (e.g., using the Spanish “perro” instead of the French “chien”). Both experiments showed more L3 than L1 intrusions during L2 naming. Furthermore, using two different tasks, we assessed if this cross-language interference was related to language inhibition. Both experiments suggested that trilinguals inhibited their L1 more strongly than their L3. Together, this suggests that a trilingual’s non-native language might experience more interference from another non-native language than from their L1, possibly because trilinguals apply more inhibition over their L1. Introduction While most research on multilingualism focuses on a bilingual’s first (L1) and second language (L2), a large proportion of the world population can hold a conversation in more than two languages (e.g., European Commission, Special Eurobarometer 386, 2012). Studies looking at third language (L3) acquisition suggest that while learning a new language, there might be more interference from the other non-native language (L2) than from the native language (L1), despite the L1 being more proficient (e.g., Puig-Mayenco, Gonz´ alez Alonso, and Rothman, 2020). However, it remains largely unknown how the languages of a trilingual influence each other once fluency is reached in all three languages (i.e., after initial stages of language learning). Across two experiments, this study therefore investigated how a trilingual’s native and non-native languages interact and compete with each other during language production. Interactions between non-native languages Research looking at interactions between native and multiple nonnative languages has mainly focused on people acquiring a new language (L3). Despite the L1 having a higher proficiency than the L2, several studies have suggested L3 acquirers might be influenced more (e. g., using syntactic structures from a known language in the L3) by the L2 than L1 (e.g., Bardel & Falk, 2007; Falk & Bardel, 2011; Rothman & Cabrelli Amaro, 2010). In a recent systematic review, Puig-Mayenco et al. (2020) examined the influence from the L1 versus L2 on third language acquisition across 71 studies, focusing on morphosyntax. Twenty of the reviewed studies showed influence exclusively from the L2 while this was the case for the L1 in only ten studies. While there are many variables that could have an impact on L3 acquisition (with typological proximity being one of the key factors), this suggests that L3 acquisition might experience more influence from another non-native language (L2) than from the native L1. While morphosyntax has been the focus of much of this research, there is also some experimental evidence showing that competition between non-native languages is also present at the lexical level. Mickan, McQueen, and Lemh¨ ofer (2020) asked Dutch-English bilinguals to learn a set of Spanish words. The next day, participants were asked to name the same pictures in either Dutch or in English. Participants then * Corresponding autor at: Department of Psychology, University of York, York YO10 5DD, UK. E-mail address: [email protected] (A. de Bruin). Contents lists available at ScienceDirect Journal of Memory and Language journal homepage: www.elsevier.com/locate/jml https://doi.org/10.1016/j.jml.2022.104386 Received 11 April 2022; Received in revised form 26 October 2022; Accepted 28 October 2022 Journal of Memory and Language 128 (2023) 104386 2 completed another naming task to assess Spanish recall immediately after the Dutch/English naming task and one week later. Spanish naming (in terms of accuracy and reaction times) was affected negatively by Dutch and English naming. However, this influence was largest for the words that had been previously named in English (L2). This suggests that at the lexical level too, a non-native language might experience more influence from another non-native language than from the native language. One explanation for these findings during L3 acquisition has been given in the form of the “L2 status” hypothesis (e.g., Bardel & Falk, 2007; 2012), which argues that the L2 and L3 are more cognitively similar than the L1 and L2/L3. Trilinguals might acquire the L2 and L3 in similar environments (e.g., in a classroom as opposed to at home or in the community) and in similar life stages (e.g., in later childhood or adulthood as opposed to from birth). The L1 and L2/L3 might also differ in their reliance on procedural versus declarative memory processes, with the latter potentially being more important for languages acquired later in life (e.g., Bardel & S´ anchez, 2017). Together, these similarities in the way non-native languages are acquired (and potentially used, if language use is restricted to e.g., the classroom) could explain why the L2 might influence L3 acquisition more than the L1. Most of this research has focused on L3 acquisition, however, and leaves open the question how the languages of a trilingual interact once a certain level of fluency is achieved in all languages. There is some anecdotal evidence to suggest that interference continues to exist between the non-native languages beyond initial stages of acquisition. For example, Williams and Hammarberg (1998) report a case study with an English-German-Swedish trilingual who, when switching out of their L3 Swedish, switched more often to their L2 German than to their L1 English. More recently, Tomoschuk, Duyck, Hartsuiker, Ferreira, and Gollan (2021) provided experimental evidence showing similar patterns, with more interference between non-native languages than between a native and non-native language. In their Experiment 1, they asked Dutch-English-French trilinguals (with a high proficiency in their L1 and L2 but a lower proficiency in their L3) to complete a phoneme detection task. Participants had to indicate whether a phoneme was present in the word corresponding to a picture (e.g., /g/ when presented with a picture of a “girl”). Importantly, participants were also presented with phonemes that were not part of the word in the target language but that were part of the translation equivalent (e.g., in the case of a picture of a girl, the phoneme /m/ is present in the L1 translation equivalent “meisje”). If there is more interference from a native language, L3 blocks should see more false alarms from the L1 (i.e., saying that the phoneme is present in the target word because it is present in the L1 translation equivalent) than from the L2. However, if there is more interference between non-native languages, there should be more L2 than L1 false alarms. The latter pattern was found, with more false alarms from L2 than L1 phonemes during the L3 task, suggesting that the L3 experienced more interference from the L2 than from the native L1. Language competition and inhibition Tomoschuk et al. (2021) suggest that the increased interference from the L2 might be related to inhibition, with trilinguals inhibiting their L1 more strongly or successfully than their L2. This interpretation is based on Green’s Inhibitory Control Hypothesis (1998), which posits that bilinguals use inhibition to avoid interference from the non-target language(s) to allow for successful production in the target language. Importantly, this hypothesis argues that the amount of inhibition applied is relative to the proficiency of the language, with bi-/trilinguals suppressing more proficient languages more strongly than less proficient languages. Evidence that bilinguals might inhibit their L1 (more so than their L2) has been found in a range of paradigms and techniques. Language switching studies (e.g., Meuter & Allport, 1999) have suggested that unbalanced bilinguals (with a higher proficiency in their L1 than L2) can show asymmetrical switch costs, with larger costs when switching back to their L1 than L2. This might be the consequence of bilinguals applying relatively large amounts of L1 inhibition during L2 naming, thus requiring more time to release this inhibition when switching back to the L1, leading to larger switching costs. The finding that more inhibition is applied over the L1 when switching to the L2 than vice versa is also supported by neuroimaging studies showing increased activation in brain regions and ERP markers associated with inhibition when switching to an L2 (i.e., the moment L1 inhibition would need to be applied) than when switching to an L1 (i.e., when less (L2) inhibition is needed; de Bruin, Roelofs, Dijkstra, & FitzPatrick, 2014; Jackson, Swainson, Cunnington, & Jackson, 2001). Other studies not using language-switching tasks have also suggested that L1 words are less accessible after L2 naming due to L1 inhibition during L2 naming. For example, Misra, Guo, Bobb, and Kroll (2012) assessed potential effects of repetition priming (i.e., faster naming when items have been named before) when naming in the L2 after the L1 and in the L1 after the L2. As expected, repetition priming was observed when the L2 was used after the L1. However, using the L1 second showed no behavioural repetition benefit and an increased N2 (an ERP component associated with topdown language control), suggesting that the L1 was suppressed during L2 naming. Interference and inhibition have also been studied through paradigms eliciting language intrusions. For example, Gollan, Schotter, Gomez, Murillo, and Rayner (2014) asked Spanish-English bilinguals to read paragraphs in one language or mixing both languages aloud. In the texts with both Spanish and English words, participants made crosslanguage intrusions (e.g., saying an English word instead of the Spanish word), in particular when trying to read words in the more dominant language. This again suggests that the more dominant/proficient language might be suppressed and consequently less accessible and more prone to intrusions. Other paradigms too have suggested that crosslanguage intrusions during production are sensitive to language control and interference (e.g., Declerck, Grainger, & Hartsuiker, 2021; Zheng, Roelofs, & Lemh¨ ofer, 2020), although they do not always reveal differences between languages (e.g., Declerck et al., 2021). Levy, McVeigh, Marful, & Anderson (2007) also examined the accessibility of L1 versus L2 word forms, by using a retrieval-induced forgetting paradigm, a task we used in Experiment 1 too and which we will refer to as the “rhyme task”. In Levy et al.’s study, EnglishSpanish bilinguals were asked to generate English rhyme words. Participants saw a probe (e.g., “spoon”) and had to generate one word that rhymed with the probe (e.g., “moon”). Prior to the rhyme task, participants named pictures in English or in Spanish. The names of those pictures could be used as rhyme words in the rhyme task (e.g., participants named a picture of a moon, with “spoon” being the probe in the rhyme task). Naming those pictures in English increased the chance of using those words in the English rhyme task (i.e., picture names produced in English were more likely to be used as rhyme responses if those words had been repeated more often in the picture-naming task). In contrast, the reverse pattern was found for words produced in Spanish (L2): the more often a picture had been named in Spanish prior to the rhyme task, the less likely participants were to use the English (L1) translation equivalent in the rhyme task. This suggests that participants had suppressed the L1 equivalents during L2 naming, thus making those L1 translation equivalents less accessible in the following rhyme task (but cf. Runnqvist & Costa, 2012 for diverging findings). Current research Following these findings, unbalanced trilinguals might suppress their more proficient native language more strongly than their less proficient non-native languages. This, in turn, could be predicted to reduce interference from a native language. This prediction is also supported by studies suggesting that L3 learning is more successful through L1 instruction than through L2 instruction, potentially because of better A. de Bruin et al. Journal of Memory and Language 128 (2023) 104386 3 regulation and inhibition of the L1 (e.g., Bogulski, Bice, & Kroll, 2019; Hirosh & Degani, 2021). However, while there is some evidence to suggest that trilinguals might experience more interference between non-native languages than from a native language (Tomoschuk et al., 2021), most research is based on L3 acquisition. Furthermore, Tomoschuk et al. (2021) showed interference in a phoneme detection task but it is unknown how potential interference between non-native languages can actually influence language production. Lastly, while L1 inhibition has been proposed as a potential underlying mechanism to explain reduced L1 interference relative to L2/L3 interference (Tomoschuk et al., 2021), the role of language inhibition has not been assessed directly and we do not know if and how language inhibition explains potential interference between languages during trilingual production. The current research therefore firstly aimed to examine how the two non-native languages of a trilingual interact with each other beyond initial stages of language acquisition, when trilinguals have already achieved an intermediate proficiency in their non-native languages. Specifically, we addressed interference between languages in the form of cross-language intrusions, such as saying the Spanish word “caballo” when the Basque “zaldi” is needed. Second, we examined why trilinguals might experience interference between non-native languages by assessing the role of L1 versus L3 inhibition. Across two experiments with trilinguals with different language combinations, we therefore used a speeded picture-naming task to study interference from the L1 versus L3 on the L2 during trilingual language production. We also examined how trilinguals inhibited their L1 versus L3 during L2 production through a rhyme task (Experiment 1) and an n-2 switching task (Experiment 2). Experiment 1 Introduction To assess language interference, Experiment 1 asked participants to complete a language-switching task in three languages (Spanish-L1, Basque-L2, and English-L3). We were especially interested in eliciting language intrusions (e.g., using a Spanish or English word when a Basque word was intended). Cross-language intrusions can be a more direct and concrete measure of language interference as they reveal noticeable mistakes in trilingual language selection. As such, this allowed us to study which language (L1 or L3) interfered more with L2 production. We therefore developed a speeded picture-naming paradigm in which participants had to alternate languages in a task presenting each picture for less than one second. If trilinguals experience more interference between non-native languages, despite a higher proficiency level in their native language, we hypothesised that they should produce more L3 than L1 intrusions during L2 target naming. As a second aim, we wanted to examine the role of inhibition during trilingual language control. To this end, we adjusted the retrievalinduced forgetting paradigm (rhyme task) used by Levy et al. (2007). In this task, participants generated a rhyme word in response to a probe (e.g., “sheep” in response to the probe “jeep”). They completed this task in Spanish (L1) and in English (L3) before and after a trilingual naming task (a different naming task than the speeded switching task). During the trilingual naming task, they named some pictures in their L1, some in their L2, and some in their L3. During the rhyme task, the probe rhymed with words used in the naming task (within-language trials) or their translation equivalents (across-language trials, see Fig. 1 and the Methods section for further information). For example, participants could see the English probe “cake” in the rhyme task preand postnaming. During the naming task, they would be asked to name pictures of a snake in English, thus increasing the likelihood of participants saying “snake” in the post-naming rhyme task (compared to the baseline pre-naming rhyme task). These so-called “within-language” control trials were used for both English-L3 and Spanish-L1 (e.g., in Spanish, a rhyme probe could be “rama”, with participants using “cama” to name the picture of a bed in the naming task). We expected within-language targets (i.e., “snake” and “cama”) to be used more often in the rhyme task postthan pre-naming due to their recent use in the naming task. However, of main interest, we also included “across-language” targets. These were L1/L3 response targets to the rhyme probe that were translation equivalents of words that had just been named in the L2 in the naming task. For example, participants would need to name a picture of a glass in Basque (“edalontzi”) in the naming task and would see the probe “mass” in English (rhyming with “glass”) and “paso” in Spanish (rhyming with “vaso”, meaning glass). In the pre-naming rhyme task we did not expect differences between languages given that the words had not yet been used in the naming task. However, in the post-naming Fig. 1. Overview of the tasks in each session in Experiment 1. A. de Bruin et al. Journal of Memory and Language 128 (2023) 104386 4 rhyme task we expected language differences on the across-language targets. Specifically, if trilinguals inhibit L1 equivalents more than L3 equivalents during L2 naming, we would expect them to use those L1 equivalents less often than the L3 equivalents during the post-naming rhyme task. For example, if a trilingual names a picture of a glass in L2 (“edalontzi”) and concurrently suppresses L1 “vaso” more than L3 “glass”, we would expect them to use “vaso” less often in response to the L1 rhyme probe “paso” than they would use “glass” to the L3 probe “mass”. Experiment 1 took place in the Basque Country. Participants had a lower proficiency in their L2/L3 than L1 and acquired only their L1 from birth. However, their language environment differs from the type of trilinguals/L3 acquirers tested in previous studies. For example, in many other studies (e.g., Tomoschuk et al., 2021), the non-native languages were both acquired as classroom languages. According to the L2 status hypothesis, interference between non-native languages could be explained by cognitive similarity as a consequence of the L2 and L3 being acquired in a classroom. In the Basque Country, however, the participants’ L2 (Basque) is a community language while the participants’ L3 (English) is treated as a foreign language that is largely restricted to the classroom. If the trilinguals in Experiment 1 show more L3 than L1 intrusions, this would suggest that this interference is not purely due to the way the L2 and L3 were acquired and are used in similar (classroom) environments. Furthermore, most participants received their (primary and secondary) education in Basque (L2) or a dual Basque-Spanish system. Tomoschuk et al. (2021, Experiment 2) suggested that language of instruction can modulate the amount of non-native language interference. Interference differences between the L1 and L2 were only found if new L3 words were taught in the L1, but not when the instruction language was L2. This raises the question whether trilinguals indeed experience more interference from their non-native language in general or whether these findings are specific to the L1 being used as the language of instruction in the classroom. If trilinguals in our Experiment 1 show more L3 than L1 intrusions (despite the L3 being taught in an L2 or bilingual school environment), this would suggest that they are better at regulating interference from their native language even if the L1 was not the main/only language of instruction. In addition, although their L2 Basque had a much lower proficiency and use than their L1 Spanish, most participants acquired Basque within the first three years of life, making it less likely that the L2 was acquired mostly through declarative memory processes. According to the L2 status hypothesis (e.g., Bardel & S´ anchez, 2017), reliance on more declarative than procedural memory systems for the L2 and L3 might explain the interference between the two. However, due to the way Basque is acquired during early childhood and is a community language while English is largely used as a classroom language, the L2 and L3 in this study are unlikely to be both relying predominantly on declarative memory. Finally, typological proximity has been argued to be a key variable in language interference, with more interactions between languages that are more similar (e.g., Puig-Mayenco et al, 2020). Basque (L2) differs substantially from both the L1 and L3 in aspects such as vocabulary and morphosyntax. However, if anything, it is more similar to the L1 than L3 in terms of orthography and phonology. Thus, if our trilingual participants experience more interference from the L3 than L1, this is unlikely to be due to typological proximity between the L2 and L3. Taking into consideration this different language profile compared to previous L3 (acquisition) studies, Experiment 1 firstly aimed to assess how much interference trilinguals experience between two non-native languages as compared to between a native and non-native language using the speeded naming task. To create the largest difference between the native and non-native language (in terms of proficiency, use, mode of acquisition, etcetera), we compared L1 versus L3 intrusions during L2 production. Next, we assessed how inhibition might be involved when trilinguals manage interference between their three languages using the rhyme task. Data availability The data and analysis script (for both Experiments) are available on https://osf.io/wmehd/. The stimuli are provided in the appendices. Methods Participants The final dataset included thirty Spanish-Basque-English trilinguals (23 female, Mage =23.3; SDage =5.6). We specifically recruited participants who acquired only one language from birth (Spanish) and who had an intermediate proficiency level in Basque and English. We ensured that participants were proficient in all languages (thus focusing on trilinguals rather than L3 acquirers) but had a much lower use of and proficiency in their L2/L3 than their L1 (thus creating a clear distinction in proficiency and use between the native L1 and non-native L2 and L3). Two additional participants were tested but excluded from data analysis. One participant was excluded because they did not know over half of the English target words (assessed in a post-experiment survey). The other participant was excluded because they did not produce any rhyme words on more than half of the trials in the baseline rhyme task. The use of novel tasks (a speeded switching task to elicit intrusions and a trilingual Table 1 Summary of the objective and subjective measures of language proficiency, language exposure, and language use for Spanish, Basque, and English. Spanish Basque English Mean SD Range Mean SD Range Mean SD Range Age of Acquisition 0 0 0–0 2.9 0.8 2–6 6.1 2.0 2–11 Picture naming (0–65) 64.8 0.5 64–65 42.6 6.7 30–51 43.8 8.0 31–56 LexTALE (% words correct) 1 88.7 8.0 83–100 67.4 16.4 28–94 43.1 22.2 3–100 Interview (1–5) 2 5 0 5–5 3.3 0.4 3–4 3.0 0.6 2–4 Self-rated proficiency 2 (0–10) Speaking 9.6 0.7x 7–10 6.9 1.5 4–10 6.0 1.7 1–9 Understanding 9.7 0.7 8–10 8.3 1.3 6–10 7.0 1.5 3–9 Writing 9.3 1.0 7–10 7.3 1.7 4–10 6.3 1.9 2–9 Reading 9.4 1.2 5–10 7.9 1.7 5–10 7.0 1.7 3–9 General 9.4 0.8 8–10 7.0 1.4 4–10 6.3 1.6 2–9 %exposure (0–100) 64.3 13.3 50–90 24.6 14.3 0–40 10.4 7.4 0–30 %speaking (0–100) 74.3 16.5 50–100 19.3 14.6 0–40 15.2 12.4 0–40 1 Data from one participant missing for Basque and English. LexTALE score is calculated as the percentage of words identified correctly. 2 Data are missing for two participants. A. de Bruin et al. Journal of Memory and Language 128 (2023) 104386 5 adaptation of a rhyme task) meant that we did not have existing data on which to base a power analysis. Furthermore, the specific language profile restricted us in the number of participants we could recruit. Thirty participants was therefore set as the target size to ensure recruitment feasibility while aiming to have as much power as possible. All participants had normal or corrected-to-normal vision and no known neurological, reading, or hearing impairments. They provided written informed consent and the study was approved by the BCBL Ethics Review Board and complied with the guidelines of the Helsinki Declaration. All participants acquired Spanish as their first language, Basque as their second language, and English as their third language. They were living in a bilingual society in which both Spanish and Basque are commonly used. Participants had an intermediate proficiency in both Basque and English and predominantly used Spanish on a daily basis (see Table 1). Their language profile was assessed through a set of objective and subjective language measures in Spanish, Basque, and English when they signed up for the database (de Bruin, Carreiras, & Du˜ nabeitia, 2017). The objective proficiency measures include a 65item picture naming task, the LexTALE (a short lexical decision task; Lemh¨ ofer & Broersma, 2012) and an interview. In addition, participants provided self-ratings of language proficiency, use, and exposure. The results from these tasks and measures are reported in Table 1. Participants also completed a short survey at the end of the study asking about their language(s) of education and time spent in English-speaking countries. Most participants (18) attended a bilingual Spanish-Basque school system; 10 participants completed their education in Basque; two participants completed their education in Spanish. Most participants (19) had not spent any significant time in an English-speaking country apart from holidays. Eleven participants indicated having spent some time in an English-speaking country (M number of months = 4.2, SD =4.2). Design of the two main tasks Speeded trilingual naming task. In the speeded trilingual naming task, participants were asked to name pictures in Spanish (L1), Basque (L2), or English (L3) in response to a country flag. A speeded task (in which participants saw each picture for only 900 ms) was used to create more time pressure and thus to induce more errors. We focused on the number of Spanish (L1) versus English (L3) intrusions during non-switch trials that were supposed to be named in Basque (L2). Rhyme task. In the rhyme task, participants were asked to generate rhyme words in response to Spanish (L1) and English (L3) probes. This task was completed twice: once at the beginning of the study (as the baseline, we will refer to this as “pre-naming”) and once at the end of the study after a naming task (“post-naming”; see Fig. 1). In the intervening naming task (a different one than the speeded naming task) before the rhyme task, people named words in either L1, L2, or L3. Words named in the L1 and L3 in that task were possible rhyme responses (i.e., targets) in the L1/L3 rhyme task (e.g., “snake” had to be named in English in the naming task and “cake” was a probe in the rhyme task). We refer to this condition as “within-language”. The words named in L2 were translation equivalents of possible rhyme responses in the L1/L3 rhyme task (e.g., “duck” named in Basque in the naming task; probe “truck”; see Fig. 2). We refer to this condition as “across-language”. There were thus three within-subject independent variables in the rhyme task: Language (L1/ L3); Session (pre-naming/post-naming); Condition (words named in the same language “within-language”/words named in Basque “acrosslanguage”). Session 1: Materials We selected sixty pictures from the MultiPic database (Du˜ nabeitia et al., 2018). Eight of these pictures were used in the speeded trilingual naming task and had to be named in the three languages interchangeably. We used a small set of pictures in this task to increase activation of each word, with the overall aim of increasing competition between words and eliciting more intrusions. All sixty pictures were used in the slow naming task preceding the post-naming rhyme task (twenty named in Spanish serving as L1 withinlanguage targets in the rhyme task; twenty named in English serving as L3 within-language targets in the rhyme task, and twenty named in Basque serving as both L1 and L3 across-language targets in the rhyme task). All picture names were non-cognates and had a high frequency. The eight words used in the speeded naming task were matched on English and Spanish frequency and on number of letters and phonemes; Spanish words had more syllables than English words (see Appendix A). We did not match the Basque words given that we were focusing on the number of English (L3) versus Spanish (L1) intrusions but did match the English and Spanish words on their Levenshtein distance to Basque words (i.e., the number of letters that would need to be changed to form the Basque word). Those eight words were chosen (from the acrosslanguage items) because they differed in their onset in the three languages. This allowed us to score cross-language intrusions even if only one phoneme was produced (e.g., if the initial response was corrected after the first phoneme). For the rhyme task, we matched the different conditions (withinFig. 2. Overview of the conditions in the rhyme task. A. de Bruin et al. Journal of Memory and Language 128 (2023) 104386 6 English versus within-Spanish; across-English versus across-Spanish; within-English versus across-English; within-Spanish versus acrossSpanish) on a range of measures including target frequency, the number of potential rhyme words for each probe, and the number of alternative rhyme words with a higher frequency than the target (see Appendix A). Due to general word length differences between English and Spanish words, the English and Spanish words could not be matched in terms of number of syllables, letters, and phonemes. However, we did ensure that the Spanish-within and Spanish-across as well as the Englishwithin and English-across conditions were matched on word length. We selected target words that had a few high-frequency rhyme competitors but not too many (to make sure that target words were relatively likely to be produced without the target being the only option). Further details about the probes are given in Appendix A. Procedure The study consisted of two in-person sessions (see Fig. 1), separated by an interval of approximately one to two weeks (M =12 days). We opted to use two separate sessions to separate the baseline and postnaming rhyme tasks (to avoid participants remembering exactly which answers they gave in the baseline task) and to separate the two naming tasks. In the first session, participants started with the baseline prenaming rhyme condition. Next, they completed the speeded trilingual naming task assessing cross-language intrusions during L2 production. In the second session, participants first completed the slow trilingual naming task that was related to the rhyme task. Next, they completed the post-naming rhyme task. After the second session, participants were asked to indicate whether there were any English words they did not know before the start of the study. On average, participants indicated knowing 38 of the 40 English rhyme targets (range 36–40; data missing from three participants). Pre-naming rhyme task (baseline) In the rhyme task, participants were presented visually and aurally with a probe word. They were asked to generate a word that rhymed with the probe. For example, a participant could be presented with ‘wig’ and could respond with ‘pig’, ‘dig’, etc. Participants were instructed to say the first rhyme word that came to mind, but not to use any proper names or names of countries or places. They were given ten seconds for each probe word. The task was completed in Spanish and English in two separate language blocks, with the order of languages counterbalanced across participants. Each language included twenty within-language trials and twenty across-language trials (see “slow naming task”). The rhyme task was completed at the start of the first session as the baseline for the rhyme task at the end of the second session. Speeded trilingual naming task Participants were first familiarised with the 8 pictures and words and were asked to read each word aloud. Participants were then presented with a speeded naming task showing pictures accompanied by the Spanish, Basque, or British flag. Each picture was presented below the country flag for 900 ms and participants were instructed to name the picture in the indicated language within that time frame. A blank screen was shown for one second before the next picture was presented. We recorded the full 1900 ms and scored responses that were given during the entire interval, even though participants were instructed to name the picture while the picture remained on the screen. In total, participants named 720 experimental trials (240 Basque non-switch trials, 160 Basque switch trials, 40 English non-switch trials, 120 English switch trials, 40 Spanish non-switch trials, and 120 Spanish switch trials). Participants named an additional ten trials that were preceded by a break and were not included in the analysis. We focused on the 240 Basque non-switch trials to examine the number of L1 (Spanish) versus L3 (English) intrusions. Crucially, these L2 non-switch trials were always preceded by another L2 trial, thus removing the immediate influence of just having used another language. L2 switch trials were preceded an equal number of times by each of the languages (e.g., 80 Basque switch trials were preceded by a Spanish trial and 80 Basque switch trials were preceded by an English trial). L1 and L3 switch trials were preceded 80 times by a Basque trial and 40 times by an L3 or L1 trial respectively. Session 2: Slow naming task: The slow naming task at the start of the second session was used to put rhyme targets in the “within-language” or “across-language” condition for the post-naming rhyme task. Before starting the naming task, participants were exposed to all sixty pictures and their names in the three languages to ensure participants were familiar with the responses. In the naming task, each picture had to be named eight times and was shown on the screen for 2 s below the country flag. Twenty of these pictures always had to be named in Basque (L2 – across-language trials for the L1 and L3 rhyme task), twenty always had to be named in English (L3 – within-language trials for the L3 rhyme task), and twenty always had to be named in Spanish (L1 – within-language trials for the L1 rhyme task). The three languages had to be used interchangeably to increase competition between the languages. There were 488 trials (8 trials preceded by a break; the 480 trials were distributed equally across languages and switch type). Each picture was presented four times as a switch trial (twice preceded by each of the two languages) and four times as a non-switch trial. Post-naming rhyme task: The rhyme task from the start of the first session was completed again, using the same probes and structure. Analysis The data are available on https://osf.io/wmehd/ and were analysed using generalised linear mixed-effects models using lme4 package version 1.1–21 in R 3.6.1. Speeded trilingual naming task. In the speeded trilingual naming task, we scored accuracy on each trial as A) no response, B) a cross-language intrusion (e.g., English instead of Basque word), C) a within-language intrusion (e.g., ‘cloud’ instead of ‘moon’), or D) another response that was not the intended target but that had a similar meaning (e.g., ‘pared’ instead of ‘muro’, with both being Spanish words for ‘wall’). We only scored the first response. For example, if multiple cross-language intrusions were made (e.g., English and then Spanish where a Basque word was required), the first was scored. Similarly, a response counted as an intrusion regardless of how much of the intrusion was produced (e.g., if just the first phoneme of the intrusion was produced and then corrected, it still counted as an intrusion). We were mainly interested in the number of L1 Spanish versus L3 English intrusions produced during L2 Basque non-switch trials. Our main analysis therefore only included L2 non-switch trials. Using generalised linear mixed-effect models (glmer; participant and item intercepts included as random effects), we compared the number of L1 versus L3 intrusions by using number of cross-language intrusions (i.e., accuracy type B) as the dependent variable (DV) and the language of intrusion (Spanish =-0.5; English =0.5) as the fixed effect. We used the “poisson” distribution within the glmer, which is suitable for count data (number of intrusions in Spanish or English) and is based on the underlying data being dichotomous (an intrusion happening or not). We also examined the same question for L2 switch trials, now also including A. de Bruin et al. Journal of Memory and Language 128 (2023) 104386 7 the language of the previous trial (i.e., L1-L2 or L3-L2 switch) as a fixed effect (switching from L3 =0.5; switching from L1 =-0.5). Rhyme task. In the rhyme task, we scored whether the response to the rhyme probe was the target word (i.e., the word used in the naming task) or not (i.e., a different word was produced or there was no response). The DV used was Target response (0 =no, 1 =yes). The three independent variables (IVs) were Language (L1 −0.5; L3 0.5); Session (prenaming −0.5; post-naming 0.5); and Condition (across language −0.5; within language 0.5). The model converged with intercepts for participants and items and the by-participant slope for condition. Results Speeded trilingual naming task Across the 720 trials in the speeded trilingual naming task (excluding the ten trials after the breaks), participants on average answered 70.5 % correctly (SD =13.3, range =37–91 % correct). Participants did not respond on 13.9 % of trials (SD =6.9). Cross-language intrusions (across all languages) were made on 14.2 % of all trials (SD =12.2) and withinlanguage intrusions on 1.1 % of trials (SD =0.8); 0.3 % of trials (SD = 0.6) were both crossand within-language intrusions (for example saying “apple” in Spanish when “horse” had to be named in English). On 0.6 % of trials (SD =0.9), a correct response was given that was not the intended target word (e.g., “muro” instead of “pared”). All participants produced cross-language intrusions, the error type of interest. We focused on the number of cross-language intrusions during Basque (L2) non-switch trials (see Fig. 3). During these trials, participants produced significantly more L3 (M =8.2 % of L2 non-switch trials showed an L3 intrusion, SD =8.6 %) than L1 (M =5.7 % of L2 nonswitch trials showed an L1 intrusion, SD =6.9; β =0.366, SE =0.064, z =5.719, p <0.001) intrusions. Basque (L2) switch trials showed similar findings. Participants produced more L3 (M =9.2 %, SD =8.6) than L1 (M =6.5 %, SD =6.8) intrusions (β =0.344, SE =0.073, z =4.695, p <0.001). There was no significant effect of the language used on the previous trial (β =0.012, SE =0.073, z =0.160, p =0.873) nor a significant interaction between language of previous trial and language of intrusion (β =0.267, SE = 0.147, z =1.818, p =0.069). Participants not only made more L3 (M = 9.9 %, SD =8.4) than L1 (M =6.1 %, SD =7.1) intrusions when an L2 trial was preceded by an L3 trial but also when it was preceded by an L1 trial (L3 intrusions M =8.5 %, SD =9.3; L1 intrusions M =6.9 %, SD = 7.1). Given the lower number of trials to be named in L3 (160 across trial types) or in L1 (160 across trial types), and given our focus on L2 production, we did not further analyse the number of intrusions during L1 or L3 trials. However, numerically more L3 intrusions were made during L1 trials (M =5.3 %, SD =7.2) than L1 intrusions during L3 trials (M = 4.0 %, SD =6.9). Given that the task required participants to use the L2 more than the other languages, the largest number of intrusions during L1/L3 trials came from the L2 (during L1 trials M =8.2 %, SD =5.7; during L3 trials M =9.7 %, SD =7.2). In summary, the speeded naming task shows that trilinguals were more likely to make L3 than L1 intrusions when having to name pictures in L2. Rhyme task The rhyme task included Session (pre-/post-naming), Language (L1/ L3), and Condition (across-/within-language naming) as the variables. Although the naming task separating the preand post-naming rhyme tasks was not of main interest, we examined accuracy to make sure participants named the pictures correctly. Accuracy was high in all three languages (Spanish M =96.6 %, SD =4.0; Basque M =89.5 %, SD =7.7; English M =89.9 %, SD =8.3). The majority of errors (M =6.2 % of all trials; SD =4.3) were no or late responses; cross-language intrusions (M =0.6 % of trials, SD =0.9) and non-target word choice (e.g., “pared” instead of “muro”, M =1.2 % of trials, SD =1.5) were rare. This confirms that participants used the target words in the naming task as intended. Fig. 4 shows the results from the rhyme task. Participants produced significantly more rhyme targets in the post-naming task (M =36.4 %, SD =7.5) than in the baseline pre-naming rhyme task (M =20.0 %, SD =5.5; β =1.052, SE =0.078, z =13.561, p <0.001). There were no main effects of language (β =0.145, SE =0.307, z =0.472, p =0.637) or condition (β =0.306, SE =0.313, z =0.978, p =0.328). These two variables did not interact with each other (β =-0.841, SE =0.615, z = -1.368, p =0.171) but importantly they did interact with session. The interaction between session and language (β =0.701, SE =0.155, z = 4.537, p <0.001) reflected that the increase between preand postnaming was larger for L3 (English) rhyme responses (Mpre =18.3 %, SD =6.0; Mpost =40.3 %, SD =9.7) than for L1 (Spanish) rhyme responses (Mpre =21.8 %, SD =7.2; Mpost =32.5 %, SD =7.4). Session also interacted with condition (β =0.421, SE =0.155, z =2.723, p = Fig. 3. Boxplots showing the percentage of language intrusions (L1 versus L3) during L2 non-switch trials. Each black square shows an individual participant (jittered). The horizontal line shows the median while the black triangle shows the mean. Fig. 4. Boxplots showing the percentage of target rhyme responses in the pre-naming task (left, baseline) and the post-naming task (right). Within each plot, the left panel represents the ‘across-language’ condition (targets named in L2 Basque) and the right panel the ‘within-language’ condition (targets named in Spanish/English, i.e., in the same language as the rhyme task). Each black dot shows an individual participant. The horizontal line shows the median while the white square shows the mean. A. de Bruin et al. Journal of Memory and Language 128 (2023) 104386 8 0.006), reflecting that the increase between preand post-naming was larger for within-language trials (i.e., trials previously named in the same language as the rhyme task; Mpre =19.9 %, SD =6.3; Mpost = 39.9 %, SD =8.5) than for across-language trials (i.e., trials previously named in L2; Mpre =20.2 %, SD =7.3; Mpost =32.8 %, SD =9.7). Importantly, there was a three-way interaction between language, session, and condition (β =-0.744, SE =0.309, z =-2.408, p =0.016). As Fig. 4 shows, there was a comparable increase relative to the baseline in target rhyme responses for L1 and L3 within-language trials. However, there was a larger increase compared to baseline in L3 across-language than L1 across-language trials, suggesting that L1 across-language targets were less accessible than L3 across-language targets. Follow-up analyses for the preand post-task separately first showed that during the pre-naming rhyme task (i.e., baseline), crucially, there was no interaction between language and condition (β =-0.230, SE = 0.653, z =-0.353, p =0.724). This was expected given that the across-/ within-language condition was only introduced in the naming task and was therefore irrelevant in the pre-naming baseline. The post-naming rhyme task, however, did show an interaction between language and condition (β =-1.208, SE =0.611, z =-1.978, p =0.048). While there was a similar number of L3 acrossand within-language rhyme targets (β =-0.089, SE =0.412, z =-0.215, p =0.830), there were more L1 withinthan across-language rhyme targets (β =1.119, SE =0.453, z = 2.467, p =0.014). This suggests that while L3 equivalents previously named in L2 were not less accessible than those named in the L3 itself, L1 translation equivalents previously named in L2 were less accessible than words named in the L1 itself. Follow-up analyses for the two conditions separately first showed an increase in target responses postcompared to pre-naming in the withinlanguage condition (β =1.247, SE =0.107, z =11.659, p <0.001). Crucially, this did not interact with language (β =0.325, SE =0.213, z = 1.527, p =0.127), demonstrating that the increase in target responses after using those words in the L1/L3 naming task did not differ for the L1 and L3. In contrast, the across-language trials showed an interaction between session and language (β =1.090, SE =0.225, z =4.847, p < 0.001), reflecting a larger increase in across-language targets for the L3 than L1. In the L3, there was a significant increase in across-language rhyme targets between the preand post-naming task (β =1.395, SE =0.156, z =8.922, p <0.001). In the L1 there was only a small increase in across-language rhyme targets that did not reach significance (β = 0.310, SE =0.162, z =1.915, p =0.056). To summarise the findings from the rhyme task, all conditions showed a similar number of target responses in the baseline measure. After the slow naming task, there was a significant increase in target responses. For words previously named in the actual language (withinlanguage), this increase did not differ for the L1 and L3. However, for words previously named in the L2 (across-language), this increase was larger for the L3 than L1, suggesting L1 translation equivalents were less accessible than L3 equivalents. Correlation between intrusions and rhyme targets We also assessed whether there was a correlation between the relative number of L1 intrusions during L2 non-switch trials (compared to L3 intrusions) and the number of L1 across-language post-naming rhyme words produced (relative to L3). In other words, we aimed to assess here whether the measure of inhibition (rhyme task) was related to the intrusions made. We were specifically interested in assessing whether people who made relatively few L1 intrusions (as compared to L3 intrusions) also produced relatively few L1 rhyme targets (as compared to L3 targets). To do this, we used the percentage difference between L1 and L3 intrusions and the percentage difference in rhyme target words between L1 and L3 in the across-language condition post-naming. There was no significant correlation between the intrusions and rhyme targets (r =-0.156, p =0.410; see Supplementary Fig. 1). As an additional posthoc check, we also examined the correlation while taking into consideration performance during the baseline rhyme task (to exclude any individual differences in how likely participants were to use target words before the naming manipulation). The correlation was not observed either when we computed the L1-L3 difference on the postnaming rhyme task relative to the pre-naming rhyme task (r =-0.225, p =0.232; see Supplementary Fig. 1). In both cases, the direction of the correlation was the opposite of expected (although not significant). This is largely driven by one outlier (see Supplementary Fig. 1) who produced a very large number of L3 intrusions. Discussion Experiment 1 had two aims. First, we examined the amount of language interference (in the form of cross-language intrusions) stemming from the native language (L1) versus from a weaker non-native language (L3) during L2 production. During a speeded picture-naming task participants showed more L3 than L1 intrusions during L2 production. This suggests that, despite the L1 being far more proficient, there was more interference from the non-native language than from the native language. Second, we aimed to examine whether trilinguals suppressed the L1 more strongly than the L3 during L2 production. Using a rhyme task after a picture-naming task showed that L1 translation equivalents were used less often than L3 translation equivalents, possibly because they were suppressed more strongly during L2 naming and consequently less accessible. The larger amount of L3 than L1 intrusions during L2 use is consistent with the L3 acquisition literature (e.g., Bardel & Falk, 2007; Falk & Bardel, 2011; Mickan et al., 2020; Puig-Mayenco et al., 2020; Rothman & Cabrelli Amaro, 2010), which shows that a non-native language has a larger influence than the native language while acquiring a new language. It is also consistent with the limited amount of research (Tomoschuk et al., 2021) suggesting that this non-native language interference might persist even after the initial stages of L3 acquisition. Here we show for the first time that increased L3 interference can disrupt L2 production in trilinguals who acquired all three languages early in life and who have an intermediate proficiency in both of their non-native languages. Furthermore, most of the acquisition literature has focused on interference stemming from a non-native language (L2) acquired before the other non-native language (L3). Tomoschuk et al. (2021) only observed increased non-native interference on an L3 but not on a (more dominant) L2. Here we show that this interference between non-native languages can also influence the non-native language that was acquired first (L2), potentially provided that this L2 has not reached very high levels of proficiency or use. One potential mechanism leading to increased interference between non-native languages might be that trilinguals apply more inhibition (or apply inhibition more successfully or efficiently) over their native language. The interpretation that trilinguals suppress their L1 most strongly is in line with Green’s (1998) inhibitory control hypothesis arguing that the amount of inhibition applied is relative to the proficiency in that language. As a consequence of the increased use of inhibition, interference from the native language might be reduced when using a less proficient language. Indeed, our trilingual participants had easier access in the rhyme task to L3 translation equivalents than L1 equivalents after L2 use, suggesting that they suppressed the L1 more strongly than the L3. These language differences were not observed in the baseline task, suggesting that they were not due to differences between the L1 and L3 target words or between the probe-target relationships. They were not observed on withinlanguage trials that were named in the rhyme language either. Multilinguals might apply language control not only by inhibiting a nontarget language but also by over-activating the less proficient languages (L2/L3, Philipp, Gade, & Koch, 2007). If the observed acrosslanguage L1/L3 differences in the rhyme task were due to L3 words being activated more strongly than L1 words, language differences should occur on the within-language trials too. The finding that only A. de Bruin et al. Journal of Memory and Language 128 (2023) 104386 9 across-language trials showed a language difference suggests that this language effect was related to reduced L1 accessibility rather than increased L3 accessibility. Various other variables have been suggested to explain increased interference between non-native languages. The L2 and L3 might be more cognitively similar (e.g., Bardel & Falk, 2007) if they are acquired in similar circumstances. In Experiment 1, however, only the L3 was acquired as a true “classroom” language. The L2 was, on average, acquired during the first three years of life and is furthermore part of the bilingual society formed by the Basque Country. Furthermore, the suggestion that non-native language interference might be related to the non-native languages being taught in the L1 (Tomoschuk et al., 2021) does not hold here considering that most participants attended education in L2 only or in a combined L1/L2 system and used Basque across their educational programme rather than as a classroom language taught through another language. Typological proximity too is unlikely to explain the results considering that the L2 differed substantially from both the L1 and L3 but, if anything, shows more phonological and orthographic overlap with the L1. Experiment 1 thus suggests that the interference between non-native languages might not be related to cognitive similarity between the L2 and L3 but rather to the way trilinguals apply more L1 than L3 inhibition. However, further analyses showed no correlation between the relative number of L1 intrusions and the relative accessibility of L1 words in the rhyme task. This could suggest that despite the same group of participants showing fewer L1 intrusions in the speeded naming task and reduced L1 access in the rhyme task, there is no direct relationship between the two findings. However, the two tasks are also very different in many other aspects (e.g., the speeded switching task requires naming of specific words while the rhyme task allows for free retrieval; the speeded switching task is a trilingual environment while the rhyme task was completed in single-language blocks). Furthermore, the rhyme task might elicit other individual differences beyond language production, including differences in the way people generate rhymes and the size of their vocabulary (i.e., the number of competitors for the target rhyme words). These many task-related differences could mask correlations within a relatively small sample. Therefore, while the rhyme task in Experiment 1 suggested that L1 translation equivalents were less accessible as a consequence of inhibition, we did want to investigate the potential role of inhibition and the relationship with intrusions further. In Experiment 2, we therefore firstly aimed to replicate the intrusion findings from the speeded naming task (in a different type of trilinguals) and we further examined the second question regarding L1 versus L3 inhibition. We used a different task (n-2 switching task, as explained below) to provide a more complete picture of the potential role of language inhibition. Experiment 2 Introduction N-2 switching task Experiment 2 used the n-2 switching task to further examine the question of inhibition of the L1 versus L3. In this task, participants are asked to switch languages on every trial. The n-2 trial (i.e., two trials before the target) either has to be produced in the same language or in a different language than the current trial (e.g., L1-L2-L1 would be an n-2 repetition trial while L3-L2-L1 would be a switch trial). If the language used on trial n-2 has to be suppressed when switching to the L2 (trial n1), participants should need more time to switch back to that language on trial n. In other words, if language X is suppressed when using the L2, it should take trilinguals more time to switch back to language X again (Lx-L2-Lx) than when a different language was used two trials ago (LzL2-Lx). Indeed, several trilingual switching studies using a range of language combinations have shown these n-2 costs, with trilinguals responding more slowly in sequences in which trial n and n-2 have to be named in the same language (repetition) than when they are named in different languages (e.g., Declerck & Philipp, 2018; Declerck, Thoma, Koch, & Philipp, 2015; Guo, Liu, Chen, & Li, 2013; Philipp et al., 2007). These repetition costs are taken as a reflection of persisting inhibition of a previously used language. Importantly, if the L1 is suppressed more strongly than the L3 when using the L2, this n2-repetition cost should be larger for the L1 than for the L3: the difference between (L1-L2-L1) and (L3-L2-L1) should be larger than between (L3-L2-L3) and (L1-L2-L3). Some studies have indeed suggested that n-2 repetition costs are largest for the trilinguals’ most dominant language(s) (e.g., Declerck et al., 2015; Philipp et al., 2007). Nevertheless, it should be noted that several studies have not shown larger repetition costs for the L1 than L2/L3 or for the L2 than L3 (e.g., Philipp & Koch, 2009; see Declerck & Koch, 2022, for a review showing inconsistencies across studies). Importantly, however, n-2 repetition effects (and potential language differences) are more likely to be explained by inhibition accounts than by overactivation of the target language. The latter would predict repetition priming rather than a cost (i.e., if the Lx is over-activated on trial n-2 and not suppressed at all, trilinguals should be faster on Lx-L2-Lx trials than on Lz-L2-Lx trials). Examining L1 versus L3 repetition costs thus allowed us to examine potential differences between L1 and L3 suppression. Importantly, we always kept the L2 as the middle trial (n-1) to specifically examine L1 versus L3 inhibition applied while switching to the L2 (contrary to previous studies, in which the middle trial for L1 repetition costs could be either L2 or L3 while the middle trial for L3 costs could be either L2 or L1). If more L1 inhibition is applied than L3 inhibition during L2 naming, we would expect L1 repetition costs to be larger than L3 repetition costs. The use of an n-2 switching task allowed us to examine suppression of L1 versus L3 through a different type of task that, in both the language-switching as well as the task-switching literature, is frequently used as a measure of dominant language/task inhibition (e.g., Declerck & Philipp, 2018; Philipp et al., 2007). This was intended to complement the, less frequently used, rhyme task from Experiment 1. Crucially, however, we designed the n-2 task differently than previous studies by always using the L2 as the n-1 trial, which allowed us to specifically look at L1 versus L3 inhibition during L2 naming. Furthermore, we assessed if and how n-2 costs were related to intrusions in the speeded naming task to elucidate the relationship between non-native language interference and inhibition. The n-2 switching task is more similar to the speeded switching task than the rhyme task, allowing us to examine this correlation in the absence of large task differences (as was the case in Experiment 1). Item-specific versus global-language control Additionally, this n-2 task allowed us to assess global inhibition (of the language as a whole) versus item-specific inhibition. When applying inhibition, multilinguals can do this reactively by just suppressing the translation equivalent of the target word in the non-target languages (e. g., when naming “caballo” in Spanish, bilinguals might just suppress the equivalent “horse” in English but no other English words). Alternatively or additionally, multilinguals might suppress all lemmas in the nontarget language (e.g., not just “horse”, but also “apple”, “dress”, etcetera). While previous studies (e.g., Philipp et al., 2007) have shown that n-2 costs can be used to study inhibition during language production, it is less clear if this inhibition is applied at a global level or is also related to specific items that have to be named. Whole-language and item-specific inhibition are not mutually exclusive: multilinguals might apply inhibition globally but also increase inhibition of the translation equivalents in an item-specific manner (e.g., Declerck & Philipp, 2017). Item-specific effects have often been studied by repeating the same pictures in a picture-naming task. For example, Misra et al. (2012) showed a facilitation effect when the same pictures had to be named in the L2 after a block of L1 naming. However, this benefit of repeating the same pictures was not present in L1 after L2 naming, suggesting inhibition of L1 forms while naming the pictures in the L2. However, this A. de Bruin et al. Journal of Memory and Language 128 (2023) 104386 16 Appendix B Spanish and English words were matched on frequency per million (set of 8: English M =66, SD =76; Spanish M =45, SD =47, t(7) = 1.592, p =0.155; set of 24: English M =50, SD =63; Spanish M =38, SD =59, t(23) =1.925, p =0.067) and number of letters (set of 8: English M =4.1, SD =1.1; Spanish M =5.1, SD =1.0, t(7) =-2.160, p =0.068; set of 24: English M =5.0, SD =1.6; Spanish M =5.5, SD =1.1, t(23) = -1.801, p =0.085). Spanish words were longer in terms of syllables (set of 8: English M =1.1, SD =0.4, Spanish M =2.1, SD =0.4, t-test could not be conducted as the variance of difference was 0; set of 24: English M =1.4, SD =0.6; Spanish M =2.4, SD =0.6, t(23) =-8.177, p <0.001) and phonemes (set of 8: English M =3.1, SD =0.4; Spanish M =4.9, SD =1.1, t(7) =-4.249, p =0.004; set of 24: English M =3.8, SD =1.2; Spanish M =5.3, SD =1.2, t(23) =-5.146, p <0.001). Spanish and English words were also matched in terms of Levenshtein Distance to the French equivalent, which was computed as the number of edits needed to change one word into the other, relative to the length of the word (with 1 meaning complete overlap and 0 meaning no overlap: set of 8: English M =0.06, SD =0.1; Spanish M =0.06, SD =0.1, t(7) =0.026, p =0.980; set of 24: English M =0.1, SD =0.1; Spanish M =0.1, SD =0.2, t(23) =0.936, p =0.359). (See Table B1). Appendix C. Supplementary material Supplementary data to this article can be found online at https://doi. org/10.1016/j.jml.2022.104386. References Anwyl-Irvine, A. L., Massoni´ e, J., Flitton, A., Kirkham, N. Z., & Evershed, J. K. (2020). Gorilla in our midst: An online behavioural experiment builder. Behavior Research Methods, 52(1), 388–407. Bardel, C., & Falk, Y. (2007). The role of the second language in third language acquisition: The case of Germanic syntax. Second Language Research, 23(4), 459–484. Table A3 (continued) Condition Language Target Probe Within English Broom Doom Within English Thief Chief Within English Bag Tag Within English Bell Cell Within English Knife Wife Within English Star Jar Within English Farm Charm Within English Snake Cake Within English Wave Cave Across English Horse Force Across English Shirt Dirt Across English Moon Spoon Across English Honey Money Across English Wheel Steel Across English Fox Ox Across English Duck Truck Across English Glass Mass Across English Wall Ball Across English Neck Wreck Across English Wing Sing Across English Snow Grow Across English Egg Peg Across English Pig Wig Across English Beach Teach Across English Path Lath Across English Cloud Proud Across English Sheep Jeep Across English Horn Torn Across English Bone Tone Within Spanish Puerta Huerta Within Spanish Pico Chico Within Spanish Flecha Brecha Within Spanish Mu˜ neca Beca Within Spanish Puente Cuente Within Spanish Conejo Espejo Within Spanish Toro Foro Within Spanish Ara˜ na Caba˜ na Within Spanish Barco Marco Within Spanish Nudo Dudo Within Spanish Trigo Digo Within Spanish Cama Rama Within Spanish Clavo Pavo Within Spanish Taza Raza Within Spanish Vaca Saca Within Spanish Silla Capilla Within Spanish Cadena Arena Within Spanish Boca Loca Within Spanish Cabeza Empieza Within Spanish Llave Nave Across Spanish Caballo Fallo Across Spanish Camisa Sonrisa Across Spanish Luna Cuna Across Spanish Miel Piel Across Spanish Rueda Pueda Across Spanish Zorro Gorro Across Spanish Pato Trato Across Spanish Vaso Paso Across Spanish Muro Duro Across Spanish Cuello Sello Across Spanish Ala Gala Across Spanish Nieve Atreve Across Spanish Huevo Llevo Across Spanish Cerdo Izquierdo Across Spanish Playa Vaya Across Spanish Camino Molino Across Spanish Nube Tuve Across Spanish Oveja Oreja Across Spanish Cuerno Tierno Across Spanish Hueso Peso Table B1 Stimuli used in Experiment 2. The first eight items (in italics) were used in the speeded naming task; all 24 items were used in the n-2 switching task. English French Spanish duck canard pato dog chien perro cheese fromage queso leg jambe pierna skirt jupe falda bed lit cama apple pomme manzana meat viande carne candle bougie vela doll poup´ ee mu˜ neca hanger cintre percha shoulder ´ epaule hombro snail escargot caracol leaf feuille hoja frog grenouille rana rabbit lapin conejo coat manteau abrigo watch montre reloj sheep mouton oveja bird oiseau p´ ajaro butterfly papillon mariposa rain pluie lluvia suitcase valise maleta city ville ciudad A. de Bruin et al. Journal of Memory and Language 128 (2023) 104386 17 Bardel, C., & Falk, Y. (2012). The L2 status factor and the declarative/procedural distinction. In J. Cabrelli Amaro, S. Flynn, & J. Rothman (Eds.), Third language acquisition in adulthood (pp. 61–78). Amsterdam: John Benjamins. Bardel, C., & S´ anchez, L. (2017). The L2 status factor hypothesis revisited: The role of metalinguistic knowledge, working memory, attention and noticing in third language learning. In T. Angelovska, & A. Hahn (Eds.), L3 Syntactic Transfer: Models, new developments and implications (pp. 85–101). Amsterdam: John Benjamins. Bogulski, C. A., Bice, K., & Kroll, J. F. (2019). Bilingualism as a desirable difficulty: Advantages in word learning depend on regulation of the dominant language. Bilingualism: Language and Cognition, 22(5), 1052–1067. Branzi, F. M., Martin, C. D., Abutalebi, J., & Costa, A. (2014). The after-effects of bilingual language production. Neuropsychologia, 52, 102–116. Brysbaert, M. (2013). Lextale_FR a fast, free, and efficient test to measure language proficiency in French. Psychologica Belgica, 53(1), 23–37. De Bruin, A., Roelofs, A., Dijkstra, T., & FitzPatrick, I. (2014). Domain-general inhibition areas of the brain are involved in language switching: FMRI evidence from trilingual speakers. NeuroImage, 90, 348–359. De Bruin, A., Carreiras, M., & Du˜ nabeitia, J. A. (2017). The BEST dataset of language proficiency. Frontiers in Psychology, 8, 522. Declerck, M., Grainger, J., & Hartsuiker, R. J. (2021). Proactive language control during bilingual sentence production. International Journal of Bilingualism, 25(6), 1813–1824. Declerck, M., & Koch, I. (2022). The concept of inhibition in bilingual control. Psychological Review.Declerck, M., & Philipp, A. M. (2017). Is there lemma-based language control? The influence of language practice and language-specific item practice on asymmetrical switch costs. Language, Cognition and Neuroscience,32(4), 488-493. Declerck, M., & Philipp, A. M. (2018). Is inhibition implemented during bilingual production and comprehension? n-2 language repetition costs unchained. Language, Cognition and Neuroscience, 33(5), 608–617. Declerck, M., Thoma, A. M., Koch, I., & Philipp, A. M. (2015). Highly proficient bilinguals implement inhibition: Evidence from n-2 language repetition costs. Journal of Experimental Psychology: Learning, Memory, and Cognition, 41(6), 1911–1916. Degani, T., Kreiner, H., Ataria, H., & Khateeb, F. (2020). The impact of brief exposure to the second language on native language production: Global or item specific? Applied Psycholinguistics, 41(1), 153–183. Duchon, A., Perea, M., Sebasti´ an-Gall´ es, N., Martí, A., & Carreiras, M. (2013). EsPal: Onestop shopping for Spanish word properties. Behavior Research Methods, 45(4), 1246–1258. Du˜ nabeitia, J. A., Crepaldi, D., Meyer, A. S., New, B., Pliatsikas, C., Smolka, E., & Brysbaert, M. (2018). MultiPic: A standardized set of 750 drawings with norms for six European languages. Quarterly Journal of Experimental Psychology, 71(4), 808–816. European Commission (2012), Special Eurobarometer 386: Europeans and their Languages. [online] Available at: https://europa.eu/eurobarometer/surveys/detail/1049. Falk, Y., & Bardel, C. (2011). Object pronouns in German L3 syntax: Evidence for the L2 status factor. Second Language Research, 27(1), 59–82. Gollan, T. H., Schotter, E. R., Gomez, J., Murillo, M., & Rayner, K. (2014). Multiple levels of bilingual language control: Evidence from language intrusions in reading aloud. Psychological Science, 25(2), 585–595. Green, D. W. (1998). Mental control of the bilingual lexico-semantic system. Bilingualism: Language and Cognition, 1(2), 67–81. Green, D. W., & Abutalebi, J. (2013). Language control in bilinguals: The adaptive control hypothesis. Journal of Cognitive Psychology, 25(5), 515–530. Guo, T., Liu, F., Chen, B., & Li, S. (2013). Inhibition of non-target languages in multilingual word production: Evidence from Uighur–Chinese–English trilinguals. Acta Psychologica, 143(3), 277–283. Hirosh, Z., & Degani, T. (2021). Novel Word Learning Among Bilinguals Can Be Better Through the (Dominant) First Language Than Through the Second Language. Language Learning, 71(4), 1044–1084. Izura, C., Cuetos, F., & Brysbaert, M. (2014). Lextale-Esp: A test to rapidly and efficiently assess the Spanish vocabulary size. Psicol´ ogica, 35(1), 49–66. Jackson, G. M., Swainson, R., Cunnington, R., & Jackson, S. R. (2001). ERP correlates of executive control during repeated language switching. Bilingualism: Language and Cognition, 4(2), 169–178. Lemh¨ ofer, K., & Broersma, M. (2012). Introducing LexTALE: A quick and valid lexical test for advanced learners of English. Behavior Research Methods, 44(2), 325–343. Levy, B. J., McVeigh, N. D., Marful, A., & Anderson, M. C. (2007). Inhibiting your native language: The role of retrieval-induced forgetting during second-language acquisition. Psychological Science, 18(1), 29–34. Meuter, R. F., & Allport, A. (1999). Bilingual language switching in naming: Asymmetrical costs of language selection. Journal of Memory and Language, 40(1), 25–40. Mickan, A., McQueen, J. M., & Lemh¨ ofer, K. (2020). Between-language competition as a driving force in foreign language attrition. Cognition, 198, Article 104218. Misra, M., Guo, T., Bobb, S. C., & Kroll, J. F. (2012). When bilinguals choose a single word to speak: Electrophysiological evidence for inhibition of the native language. Journal of Memory and Language, 67(1), 224–237. Philipp, A. M., Gade, M., & Koch, I. (2007). Inhibitory processes in language switching: Evidence from switching language-defined response sets. European Journal of Cognitive Psychology, 19(3), 395–416. Philipp, A. M., & Koch, I. (2009). Inhibition in language switching: What is inhibited when switching between languages in naming tasks? Journal of Experimental Psychology: Learning, Memory, and Cognition, 35(5), 1187–1195. Protopapas, A. (2007). Check Vocal: A program to facilitate checking the accuracy and response time of vocal responses from DMDX. Behavior Research Methods, 39(4), 859–862. Puig-Mayenco, E., Gonz´ alez Alonso, J., & Rothman, J. (2020). A systematic review of transfer studies in third language acquisition. Second Language Research, 36(1), 31–64. Rothman, J., & Cabrelli Amaro, J. (2010). What variables condition syntactic transfer? A look at the L3 initial state. Second Language Research, 26(2), 189–218. Runnqvist, E., & Costa, A. (2012). Is retrieval-induced forgetting behind the bilingual disadvantage in word production? Bilingualism: Language and Cognition, 15(2), 365–377. Tomoschuk, B., Duyck, W., Hartsuiker, R. J., Ferreira, V. S., & Gollan, T. H. (2021). Language of instruction affects language interference in the third language. Bilingualism: Language and Cognition, 24(4), 707–718. Van Assche, E., Duyck, W., & Gollan, T. H. (2013). Whole-language and item-specific control in bilingual language production. Journal of Experimental Psychology: Learning, Memory, and Cognition, 39(6), 1781–1792. Williams, S., & Hammarberg, B. (1998). Language switches in L3 production: Implications for a polyglot speaking model. Applied Linguistics, 19(3), 295–333. Wilson, M. (1988). MRC psycholinguistic database: Machine-usable dictionary, version 2.00. Behavior Research Methods, Instruments, & Computers, 20(1), 6–10. Zheng, X., Roelofs, A., & Lemh¨ ofer, K. (2020). Language selection contributes to intrusion errors in speaking: Evidence from picture naming. Bilingualism: Language and Cognition, 23(4), 788–800. A. de Bruin et al.