Language balance rather than age of acquisition: A study on the cross-linguistic gender congruency effect in Portuguese– German bilinguals

Bilingualism: Language and Cognition cambridge.org/bil Research Article Cite this article: Sá-Leite, A. R., Flores, C., Eira, C., Haro, J., Comesaña, M. (2023). Language balance rather than age of acquisition: A study on the cross-linguistic gender congruency effect in Portuguese–German bilinguals. Bilingualism: Language and Cognition,1–14. https://doi.org/10.1017/S1366728923000378 Received: 8 March 2022 Revised: 22 March 2023 Accepted: 22 April 2023 Keywords: age of acquisition; bilingualism; language balance; gender congruency effect; grammatical gender; proficiency Corresponding author: Ana Rita Sá-Leite; Email: anar.saleite@gmail. com © The Author(s), 2023. Published by Cambridge University Press. This is an Open Access article, distributed under the terms of the Creative Commons AttributionNonCommercial licence (http:// creativecommons.org/licenses/by-nc/4.0), which permits non-commercial re-use, distribution, and reproduction in any medium, provided the original article is properly cited. The written permission of Cambridge University Press must be obtained prior to any commercial use. Language balance rather than age of acquisition: A study on the cross-linguistic gender congruency effect in Portuguese– German bilinguals Ana Rita Sá-Leite1,2 , Cristina Flores3, Carina Eira4, Juan Haro5 and Montserrat Comesaña6,7 1 Cognitive Processes & Behaviour Research Group, Department of Social Psychology, Basic Psychology, and Methodology, University of Santiago de Compostela, Santiago de Compostela, Spain; 2 Institut für Romanische Sprachen und Literaturen, Goethe-Universität Frankfurt, Frankfurt am Main, Germany; 3 Center for Humanities, School of Arts and Humanities, University of Minho, Braga, Portugal; 4 School of Arts and Humanities, University of Minho, Braga, Portugal; 5 Department of Psychology, Research Center for Behavior Assessment (CRAMC), Universitat Rovira i Virgili, Tarragona, Spain; 6 Psycholinguistics Research LineCIPsi, School of Psychology, University of Minho, Braga, Portugal and 7 Centro de Investigación Nebrija en Cognición (CINC), Universidad Nebrija, Madrid, Spain Abstract The cross-linguistic gender congruency effect (GCE; a facilitation on gender retrieval for translations of the same gender) is a robust phenomenon analysed almost exclusively with late bilinguals. However, it is important to ascertain whether it is modulated by age of acquisition (AoA) and language proficiency. We asked 64 early and late bilinguals of European Portuguese and German to do a forward and backward translation task. A measure of language balance was calculated through the DIALANG test. Analyses included this factor along with the gender congruency between translations, the target language, and the AoA of both languages, among others. Results showed a GCE for European Portuguese that was independent of the AoA and greater the higher the language imbalance. We propose that changes in proficiency in any of the languages create situations of dependency between them which allow cross-linguistic gender interaction to occur and effects to emerge depending on gender transparency. Introduction When communicating, bilinguals have to control or even inhibit to a certain extent the language that they do not intend to produce. Interference of the non-target over the target language must be kept minimal for accurate output to be achieved. Yet, this task can be especially tricky for certain aspects of grammar. There is one feature notorious for being problematic and creating situations of long-term interference during language processing that are not easily overcome, and that is grammatical gender (Carroll, 1989; Franceschina, 2005; Hawkins, 2009). Grammatical gender is an inherent abstract characteristic of nouns that partially determines the form of other words in speech 1 . It is present in gendered languages, which have gender systems that classify nouns according to different values. The number (and type) of gender values depends on the language itself (Corbett, 2013). For instance, in European Portuguese (EP) nouns can be either masculine (M, clock: “relógio”), or feminine (F, table: “mesa”), whereas in German nouns can also be neuter (N, room: “Zimmer”) besides feminine and masculine. Due to the abstract nature of grammatical gender, there is no particular reason for assigning one value or another to nouns, which makes gender assignment arbitrary in terms of semantics. Because of this arbitrariness, the gender value assigned to a certain noun might differ across languages. Thus, it is easy to imagine how tricky this feature can be for bilinguals, both in terms of acquisition and of processing. Indeed, even though in EP “clock”is masculine and “table”is feminine, in German this classification is reversed and hence “clock”is feminine (“Uhr”), and “table”is masculine (“Tisch”). Other nouns, however, keep the same gender (e.g., “door”, which is feminine in both EP and German, “porta”and “Tür”, respectively). This cross-linguistic (in)congruency between gender values is reflected in two specific terms: heterogeneric and homogeneric translations. Heterogeneric pairs have This article has earned badges for transparent research practices: Open Data and Open Materials. For details see the Data Availability Statement. 1 This study does not assess natural gender, a semantic-based feature that does not comply with the abstract and arbitrary nature of grammatical gender. https://doi.org/10.1017/S1366728923000378 Published online by Cambridge University Press different gender values and thus are gender incongruent, whereas homogeneric pairs have the same gender value and thus are gender congruent. This situation of (in)congruency evidently increases the difficulty of learning and correctly using the gender of nouns across multiple languages (Franceschina, 2005). Yet, the target-like assignment of gender is still crucial to assure accurate output in terms of agreement, as the form of other words such as articles or adjectives changes in order to agree with the gender of the head noun. For instance, in EP we would say “O relógio caro” (the expensive clock) but “A mesa cara”(the expensive table). The repercussions that the mismatches between gender systems can have on the acquisition, representation and retrieval of gender in bilinguals have been a focus of interest of many researchers (e.g., Egger et al., 2018; Kupisch et al., 2022;Lemhöferetal.,2008; Unsworth, 2008, among many others). In the present study, we focused on the representation and processing of grammatical gender in both early and late bilinguals. More specifically, we asked EP and German bilinguals to translate bare nouns in backward (from L2 to L1) and forward (from L1 to L2) direction in order to understand how gender values are selected during the lexical access to nouns depending on the gender congruency between equivalent translations. We did so by focusing on the so-called gender congruency effect (Sá-Leite et al., 2019,2020). Importantly, we took into consideration two relevant individual background variables that have remained poorly assessed in the literature on gender processing by adult speakers: the age of acquisition (AoA) of the two languages and the language balance, assessed through a proficiency-based measurement of each language. Tackling gender retrieval during bilingual noun production The study of the representation and processing of grammatical gender during bilingual language production has attracted the attention of many scholars during the last decades. Particularly, the debate has focused on whether the gender systems of each language are represented separately in the bilingual’s mind or rather, languages share a unique system (the autonomous vs. the integrative view, see Sá-Leite et al., 2020). To understand this debate, we must first comprehend how gender is represented in the linguistic system as a feature. It is widely consensual among theories of lexical selection that gender is located at the lexico-syntactic level of word representation in the form of one node per gender value (e.g., the M, F, and N gender nodes; Caramazza, 1997; Levelt et al., 1999). These gender nodes are connected to all the nouns in the lexicon as a function of their gender values, so that “Tisch”(“table”, M) is connected to the masculine node (see Figure 1). Yet, bilinguals could either have an autonomous gender system per language, each one with its own nodes (e.g., a masculine gender node for EP and a masculine gender node for German), or rather an integrated system in which all lexical entries belonging to the same value are connected to the same gender node, regardless of the language they belong to (e.g., “porta”and “Tür”[“door”] would be connected to the same feminine gender node). To explore this issue, researchers have relied on a specific effect, the cross-linguistic gender congruency effect (GCE) obtained mainly with two different experimental settings: picture naming and forward translation tasks. In the former, participants are asked to name pictures orally in the second language (L2, e.g., bilinguals of EP [first language, L1] and German [L2] would say “Tisch”when presented with the picture of a table); in the latter, participants have to orally translate to the L2 nouns written in their L1 (e.g., saying “Tisch”[German] when presented with the word “mesa”[EP]). In both tasks, conditions of gender congruency and incongruency are created through the selection of homogeneric and heterogeneric translation pairs. According to the integrative view of bilingual gender representation, variations on the response times (RTs) of the participants are expected depending on the type of translation pair. These variations on the RTs are not expected if each language has its own autonomous gender system with its own nodes. Imagine that a bilingual of EP and German wanted to produce the German noun “Tisch”. The processing of this noun would require first the activation and selection of its semantic features, that later would activate the grammatical and syntactic information of the word associated to that concept (tisch) as well as its morpho-phonological realization (/tisch/) 2 . Hence, the masculine gender node would receive activation coming from the semantic representation of “TISCH” and would thus be selected.2In the experimental context of a naming or translation task, when bilingual participants are presented with the image of a table or the word “table”in their L1 (e.g., in EP, “mesa”), they activate not one but two appropriate lexical entries (“Tisch”and “mesa”; e.g., Hatzidaki et al., 2011; Klaus et al., 2018). The lexico-syntactic representation of tisch would activate the masculine gender node, conversely to mesa, which would activate the feminine node. Since both gender nodes are active, they compete for selection (see Figure 2). This process of competition is mostly addressed in the WEAVER++ model of language processing by Levelt et al. (1999). According to the authors and as it has been shown in several computational simulations, for a node to be selected, activation has to reach a threshold defined by the difference of activation across nodes (the Luce ratio, see Roelofs, 1992). Hence, the greater the activation strength of a non-target node is, the harder it is for the target node to reach the difference of activation marked by the threshold for selection. Conversely, if different sources of activation converge in the same node, facilitation is observed. Yet, if gender systems were autonomous and thus separated depending on the language, convergence on the same node would not occur and facilitation would not be observed (and the same would apply for mechanisms based on competition rather than on facilitation as competition would not occur between autonomous systems). Results from the naming and translation tasks support the integrative view of the bilingual gender representation as they reveal a consistent cross-linguistic GCE by which RTs are significantly lower for the gender congruent condition (homogeneric translations) than for the gender incongruent one (heterogeneric translations; for a review, see Sá-Leite et al., 2019). In other words, in the case of homogeneric translations, the threshold for selection is being reached more easily in comparison to when activation is spread to the opposite gender. This effect has been tested almost exclusively in research on late bilinguals (but see Fuchs, 2022, for an ongoing study on Spanish heritage speakers), focusing on their L2, and observed in picture naming tasks (Bordag, 2004; Bordag & Pechmann, 2007; Klassen, 2016; Lemhöfer et al., 2008; Manolescu & Jarema, 2015; Morales 2 Note that how this selection occurs is a matter of discussion in the literature, especially when concerning monolingual and first language processing. Some authors argue that gender is selected automatically without the intervention of competitive mechanisms (Caramazza et al., 2001). Others argue that gender is selected competitively, but only in the presence of elements of agreement, for which gender selection is required to encode the form of the other words (Levelt et al., 1999). In this study, we will adopt a competitive view on gender selection that is however independent on the presence of elements of agreement, as this is thus far the only way we are capable of explaining the results on bilingualism (see Sá-Leite et al., 2019,2020 for overviews). 2 Ana Rita Sá‐Leite et al. https://doi.org/10.1017/S1366728923000378 Published online by Cambridge University Press et al., 2011; Paolieri et al., 2010; but for null results see Costa et al., 2003), as well as in translation tasks (Bordag & Pechmann, 2008; Manolescu & Jarema, 2015; Paolieri et al., 2010,2019; Salamoura & Williams, 2007) featuring multiple language pairs from the Germanic, Slavic, and Romance language families (i.e., Czech, Dutch, French, German, Greek, Italian, Romanian, Russian, and Spanish). Current gaps in the literature The cross-linguistic GCE is without a doubt highly informative when it comes to the organization of grammatical gender within the bilingual mind and appears to clearly support the integrative view. However, this scenario is likely to be more complex than what has been assumed so far. On the one hand, there are subtle differences in the outcome obtained by the above-mentioned tasks. Indeed, naming tasks seem to better detect the crosslinguistic GCE in comparison to translation tasks. This idea was first pointed out by Bordag and Pechmann (2008), who failed to obtain the effect in three translations tasks with Czech and German late bilinguals. Still, they had obtained the effect in naming tasks with participants of the same population (Bordag & Pechmann, 2007). The authors proposed that differences across tasks in the time course of gender activation were responsible Figure 1. Representation of gender selection during lexical access in German Note. Representation has been simplified as it intends to be illustrative of gender selection during noun production through the spread of activation. The conceptual stratum represents the abstract semantic features associated with each word, here illustrated through the English noun “table”. Continuous bold lines indicate selection; discontinuous lines represent features (N) that have been neither activated nor selected. M = Masculine; F = Feminine; N = Neuter. Figure based mainly on the WEAVER++ model of lexical access (Levelt et al., 1999). Figure 2. Production of “table”in German in a shared gender system with Portuguese Note. Lexical access to the word “table”for a Portuguese and German bilingual (without mechanisms of inhibition or control considered). Representation has been simplified as it intends to be illustrative of gender selection and gender nodes within bilingualism during noun production. The conceptual stratum includes the abstract semantic features associated with each word, here represented by the English noun “table”. Continuous bold lines indicate selection; continuous fine lines indicate activation but not selection; discontinuous lines represent features (N) that have been neither activated nor selected. Spread of activation starts on the conceptual stratum. The masculine gender node is selected, whereas the feminine gender node is activated by the lexical representation of the word “table”in Portuguese (“mesa”), and has hence been a competitor for selection. M = Masculine; F = Feminine; N = Neuter. Figure based mainly on the WEAVER++ model of lexical access (Levelt et al., 1999). Bilingualism: Language and Cognition 3 https://doi.org/10.1017/S1366728923000378 Published online by Cambridge University Press for the absence of the effect. A recent proposal on the mechanisms underlying gender retrieval supports this idea by noting that gender effects are sensitive to the time course of lexical access because of the inherent low degree of activation of gender nodes (Sá-Leite, 2021). Indeed, in naming tasks, the activation spreads in parallel from the shared concept across languages to both L1 and L2 lemmas. So, gender nodes of one language and another are activated practically at the same time, creating an ideal situation for facilitation or competition to arise. Yet, in forward translation tasks, the word-form representation of the L1 noun that appears on the screen is activated along with its lemma earlier than that of the L2 translation equivalent. In fact, it is the activation of the L1 noun that spreads to the lemma of the L2 translation equivalent, for which the gender nodes are inevitably activated after the activation of the L1 gender node. If gender nodes accumulated high degrees of activation, this probably would not be a problem, but meta-analytic research by Sá-Leite et al. (2020,2022) very much suggests that gender nodes accumulate low levels of activation, which in turn originates slippery experimental gender effects characterized by a high degree of heterogeneity and small sizes. In this sense, Sá-Leite (2021) explains that the level of activation would depend on the language itself. Languages like Italian, EP or Spanish have a high degree of transparency as more than 60% of their nouns end in the quite simple ortho-phonological gender cues “-a”(for feminine nouns) and “-o”(for masculine nouns). Many studies have shown that transparent nouns are more accurately processed and require more cognitive resources to be accessed (Caffarra et al., 2014). Following the Dual-Route model of gender retrieval by Gollan and Frost (2001), transparent nouns seem to rely on an extra route of gender selection besides the lexical memory-based route: the form-based route represented by the noun ending. This does not seem to be the case for languages like French, German, or Dutch, which rely on an extremely complex and sometimes contradictory body of gender-form regularities and are hence considered less transparent for gender (Kupisch et al., 2018). For instance, in German, Köpcke (1982) and Köpcke and Zubin (1983) found a quite high number of gender regularities in monomorphemic and extended monomorphemic nouns (44 of which are not coincident with any gender morphemes), which, in addition, depend on the case (e.g., nominative vs. dative), as gender intertwines with declension. Consequently, regularities in opaque languages are not as useful as the typical “-a”and “-o”of transparent languages and the retrieval of gender seems to rely mainly on one source: the lexical memory-based route. Sá-Leite (2021) proposes that the existence of an extra source of gender activation for most nouns in gender-transparent languages increases the resting level of activation of gender nodes. Ultimately, in these cases, gender selection involves higher levels of activation, which might make gender competition for selection (e.g., masculine vs. feminine) more easily observed. To be precise, bilinguals are the population in which genuine effects of gender congruency are most consistently obtained, even for languages with low degrees of transparency (Bordag & Pechmann, 2007; Lemhöfer et al., 2008; for a review see also Sá-Leite et al., 2019). This should not surprise us, as bilinguals have a systematic double source of gender activation due to the activation of equivalent translations. Yet, we may note that when the time course of activation of both competing nodes differs, effects may become slippery, especially when less transparent languages such as German are involved. To be precise, null effects within the area of the cross-linguistic GCE have been obtained for Germanic languages in translation tasks but not in naming tasks (Bordag & Pechmann, 2008; Salamoura & Williams, 2007). In sum, the slipperiness of the results obtained with translation tasks depending on the transparency of the language is still a matter to be explored. On the other hand, all the studies on the cross-linguistic GCE have focused on the impact of the L1 on the L2, but not of the L2 on the L1, even though there is increasing evidence suggesting that the acquisition and presence of an L2 within the linguistic system may modulate the representation and processing of the L1 (e.g., Hamann et al., 2017; Ulbrich & Ordin, 2014). More specifically, evidence is quite robust for long-term residents of the L2 environment who achieve near-native L2 proficiency (Schmid, 2009). Thus, a comprehensive view of the bilingual gender selection requires a thoughtful examination of language interference when gender is retrieved for the L1 and not only for the L2. In this sense, two variables that have been shown to be critical for language processing in bilingualism remain mainly unexplored in this area: AoA and language proficiency. Regarding the AoA, most of the studies on the cross-linguistic GCE have tested participants that acquired their L2 later in life (after the age of 10 –Bordag, 2004; Bordag & Pechmann, 2007, 2008; Klassen, 2016; Lemhöfer et al., 2008; Morales et al., 2011; Paolieri et al., 2010,2019; Salamoura & Williams, 2007). We believe this bias for late learners should be broken in order to draw more precise conclusions on the way gender is organized and selected in bilinguals. Indeed, only two published studies have tested early bilinguals (Costa et al., 2003; Manolescu & Jarema, 2015). Costa et al. (2003) conducted picture naming tasks with speakers of Croatian and Italian, Spanish and Catalan, and Italian and French that had acquired their L2 after an average age of five. However, the study used small samples ranging from 10 to 22 participants and was appointed with multiple methodological flaws (for an overview, see Sá-Leite et al., 2019; see also Lemhöfer et al., 2008). The other study that previously tested early bilinguals was that of Manolescu and Jarema (2015), who unfortunately provided scarce information regarding the linguistic background of their participants. They only stated that they were children of Romanian parents that moved to Montreal and started acquiring French during “childhood”, which is not particularly specific. In sum and due to the scarcity of research with early bilinguals, thus far conclusions on the organization of the bilingual gender system derived from research with late learners cannot be generalized to early learners. Hence there is a need to test early learners of an L2 or even simultaneous bilinguals, assuring a detailed assessment of their linguistic background, so that we can understand gender retrieval in bilingual language production in a broader sense. Finally, language proficiency is perhaps the variable that concerns us the most. As pointed out by Sá-Leite et al. (2020) in their meta-analysis on the cross-linguistic GCE, proficiency in the L2 has never been included as a factor in the analyses of any study on this effect. It has been, however, controlled in many different ways: informal interviews that took place prior to the experiment (e.g., Bordag & Pechmann, 2007), different self-informed subjective questionnaires made by the authors (e.g., Costa et al., 2003; Lemhöfer et al., 2008; Morales et al., 2011; Paolieri et al., 2010, 2019)–in some cases the authors did not report any information about what the participants were asked, which skills were considered, or what scores were obtained in average (e.g., Manolescu & Jarema, 2015)–, self-informed standardized questionnaires such as Hermans et al.’s(1998) or Bachman and Palmer’s(1989; e.g., Bordag & Pechmann, 2007; Salamoura & Williams 2007), ratings 4 Ana Rita Sá‐Leite et al. https://doi.org/10.1017/S1366728923000378 Published online by Cambridge University Press in measures such as the familiarity of the target nouns (e.g., Bordag & Pechmann, 2007), and official language tests that had been successfully completed by the participants (although it is not said when –e.g., Bordag, 2004; Salamoura & Williams, 2007). To our knowledge, only Klassen (2016) seems to have used an objective measure of proficiency (the proficiency test of the Goethe-Institut, 2010). In sum, most studies have exclusively used subjective and non-standardized ways of measuring L2 proficiency. Likewise, when checking the self-informed data on the L2 AoA, L2 age of first exposure, and L2 time of exposure across all studies, Sá-Leite et al. (2020) noticed great variations across studies for populations that were said to have the same level of proficiency. This might be another index of a subjective and nonprecise way of understanding what constitutes proficiency. It is certainly quite baffling that a variable that has been shown as extremely relevant for bilingual language processing in many other areas (see, for instance, Bultena et al., 2015; Lim & Christianson, 2015; Prior et al., 2007) has been mostly neglected in this specific area. In fact, we believe that the language proficiency of the participants should not only be properly controlled and reported in the subsequent works on bilingual gender processing but should also be included in the analyses. The literature on other areas of bilingualism shows enough evidence to suspect it may have a great role in determining the way languages interact and hence in how gender is selected in one language or another (Kupisch et al., 2013; Soares et al., 2019). More precisely, it is essential to test bilinguals’proficiency in both of their languages and to assess the balance between them, i.e., their relative language dominance. We therefore ought to (1) use more objective measurements to test proficiency and (2) include this variable in the analysis, along with gender congruency, to truly understand its role in the effect. Theoretical background for the role of AoA and proficiency in gender retrieval We will now present a proposal on how AoA and proficiency may impact grammatical gender representation and processing. We will base ourselves on the idea of relative language dominance (henceforth: language balance), which refers to the degree of balance between the languages of the bilingual speaker. To do so, we will rely on two popular and highly supported models of bilingual lexical processing: the developmental Bilingual Interactive Activation model (BIA-d, Grainger et al., 2010) and the Multilink (Dijkstra et al., 2019). The BIA-d model is an extension of the connectionist but also localist BIA model (Grainger & Dijkstra, 1992) that offers an interesting view on the adaptations that our linguistic system experiences during the acquisition of an L2. The Multilink, however, is the most recent connectionist model of bilingual language production and comprehension. By combining both models (and also considering that language learning involves an on-going fine-tuning of the learning system that continues across life span, see Ramscar et al., 2014; also Chuang & Baayen, 2021), we believe we can address the dynamics of a flexible system susceptible to changes in proficiency depending on variations in word frequency of exposure and use. Following the BIA-d model, as well as extensive empirical evidence on the matter, when the L2 proficiency increases, the L2 dependency on the L1 decreases, and hence the ability of the L1 to create interference during L2 processing decreases as well (Abutalebi & Green, 2007; Pivneva et al., 2012). Indeed, it is known that the higher the L2 proficiency of the bilingual, the better their performance during L2 production at different levels (Costa & Caramazza, 1999; Kroll et al., 2002; Pivneva et al., 2012). The tenets of the Multilink are in line with this: as the frequency of encounters with words increases, the links between their lexicosyntactic representations (i.e., lemmas) and their grammatical features become stronger. We hypothesize that this would also affect grammatical gender and, hence, the stronger the link between the L2 lemma and the gender node, the easier the retrieval of gender and the less the dependency on the L1 gender node as well as its ability to interfere. Consequently, effects of cross-linguistic gender interaction such as the GCE should be smaller, the higher the proficiency in the target language. Regarding the AoA, it is still an open question whether early bilinguals have a better language control (Berken et al., 2016, 2017; Bonfieni et al., 2019; Wattendorf et al., 2014), yet what seems clear is that other kinds of factors, such as daily exposure to each language or changes in the language environment, can modulate the interaction of both languages during language processing regardless of when the L2 was acquired (for a review, see Van Hell & Tanner, 2012; see also Bonfieni et al., 2019). This occurs largely because these factors directly influence the degree of proficiency of the speaker in each language (e.g., Dussias & Sagarra, 2007; Levy et al., 2007; but for more detail, see Van Hell & Tanner, 2012). Therefore, whether or not a cross-linguistic GCE is observable should depend mainly on the degree of proficiency in each language, and being an early bilingual should not be an impediment to language interaction happening (for a study in language comprehension that supports this prediction, see Paolieri et al., 2020). In fact, many studies assume that AoA effects may be leveled by increasing proficiency (see Gagarina & Klassert, 2018). We thus believe that modulations in proficiency of either of the two languages should produce changes in the strength of the links connecting grammatical gender to lemmas. The changes in strength create states of dependence or interference of one language over another, somehow regressing in the phases of acquisition defined by the BIA-d model (see our proposal in Figure 3). In this sense, note that both the L2 or the L1 could suffer changes in their representational state, depending on these fluctuations in proficiency, in line with previous evidence (e.g., Dussias & Sagarra, 2007; Guo et al., 2011; Linck et al., 2009; see also Morales et al., 2014). Thus, we do not expect AoA to be a better predictor of the GCE than proficiency. The present study In the present study, we aimed to: (1) explore the cross-linguistic GCE not only in an L2 but also in an L1 by assessing EP and German bilingual adult speakers to understand how crosslinguistic interference might modulate gender retrieval as a whole; (2) test bilinguals who speak languages with different degrees of transparency (a more transparent [EP] and a more opaque [German] language) to examine if the degree of gender transparency of the languages may have any impact on the crosslinguistic GCE; (3) for the first time test the role of AoA and language balance in the effect. To do so, we conducted translation tasks as is traditionally done in this field of research. However, instead of using only forward translation tasks as it has been previously done in the literature (Bordag & Pechmann, 2008; Manolescu & Jarema, 2015; Paolieri et al., 2010; Salamoura & Williams, 2007) we also used backward translation tasks, which allowed us to explore the effect of one language on another and vice-versa during language production. We recruited participants Bilingualism: Language and Cognition 5 https://doi.org/10.1017/S1366728923000378 Published online by Cambridge University Press from two different populations: (a) the so-called heritage speakers, i.e., early bilinguals of EP and German that had Portuguese parents and were born or lived during childhood in a German-speaking country (either Germany or Switzerland) with EP as home language, and (b) monolingual-raised native speakers of EP who started to learn German after the age of 10 in a classroom setting. The early bilinguals differed in the number of years they had been in German-speaking countries, with some living there even during adulthood and others leaving during childhood. At the moment of testing, 29 early bilinguals lived in Portugal, 12 in Germany and one participant in the German-speaking part of Switzerland. All early bilinguals have in common having acquired German in childhood through immersion before the age of 7 (following Hyltenstam & Abrahamsson’s[2003] cut-off of early bilinguals). Consequently, they differed in their degree of proficiency in each language, suffering certain imbalances between languages due to varying proficiency in either EP or German (from now on, L1 and L2). 3 According to our proposal, for some participants, the strength of the links between the gender nodes and the L1 lemmas should have likely decreased, creating a situation of possible interference between languages, even though they were early bilinguals and highly proficient in the L2. We hence considered not only the proficiency of the L2 but also that of the L1. More specifically, we used a standardized and highly valid and reliable measurement of the degree of proficiency in both languages, the DIALANG Vocabulary Size Placement Test (see Alderson, 2005). We obtained a measure of language balance by subtracting the DIALANG score in one language from the score in the other, thus obtaining a differential-based dominance index, as suggested by Birdsong (2015). We therefore tested the following hypotheses: (a) the crosslinguistic GCE is dependent on the language balance, so that the higher the imbalance, the greater the effect; (b) the effect can be obtained in both an L1 (EP) and an L2 (German), even though it might be modulated by the opaqueness of the Germanic language, being null for this language as obtained in previous studies with translation tasks assessing German as L2 (Bordag & Pechmann, 2008; Salamoura & Williams, 2007); (c) hence the effect does not depend on the AoA. The data and scripts used in this study are available online at the following link: http://doi.org/10.17605/OSF.IO/UE9XH Method Participants Seventy-four voluntary adult bilinguals of EP and German (62 female; M age = 38.12 years, SD = 9.73) were recruited online via email and social media and personal contact was made with each one of them. The requirements for participation for late bilinguals were (1) having started to learn German in a classroom context as teenagers or adults; and (2) having studied German for 5 years or more or having lived in a German-speaking country as adults. As for early bilinguals, the requirements were (1) being born or having immigrated to a German-speaking country before age 7 (Hyltenstam & Abrahamsson, 2003); and (2) having lived there for 6 or more years. Ten of them reported moderate to high proficiency in another gendered language apart from these languages in the Language History Questionnaire (LHQ, Li et al., 2020). Those were French, Spanish and Italian. 4 In addition to AoA, country of residence and years living in a German-speaking country, participants were also asked to selfrate their proficiency in all languages they knew and to estimate their degree of contact with each language in their daily life (by dividing 100% of contact among all relevant languages). The early bilinguals (n = 42) grew up as Portuguese-descendant second generation immigrants in Germany or in the German-speaking part of Switzerland. They started to acquire EP from birth as heritage language, i.e., as the main language spoken within the family. As is typical for heritage speakers, contact with the majority language (German) started either from birth or during pre-school age. 5 Thus, all speakers were either Figure 3. How L2 gender representation develops during acquisition following the BIA-d model Note. L1 = First language; L2 = Second language. Discontinuous lines represent weak connections. The thinner the line, the weaker the connection. In our predictions, the representational state of the linguistic system may vary depending on the proficiency of one language or another. Figure based on Grainger et al. (2010). 3 All early bilinguals acquired EP as their heritage language, which is an L1 (Flores, 2015). For the majority of these speakers German (the societal language) is an early L2, although some had at least some contact with German from birth since they were born in Germany (n = 21). For the sake of simplicity, we will refer to EP as the L1 and to German as the L2 throughout the text, especially since we treated the AoA as a continuous variable in the model of analysis (see Results section). 4 We evaluated the effect of knowledge of a gendered language (moderate to high proficiency) on our models. However, we found that this variable had no impact on the outcome. The main effect of this variable was not significant, and the significance of the other effects remained unchanged. This means that the results that were statistically significant before remained significant (all ps < .05) and the ones that were not remained unchanged (all ps > .05). Based on these findings, we determined that this variable has no influence on the results and therefore we did not include it in the models presented in the manuscript. 5 No participant from Switzerland spoke Swiss German within the family; all acquired standard German in pre-school age. 6 Ana Rita Sá‐Leite et al. https://doi.org/10.1017/S1366728923000378 Published online by Cambridge University Press simultaneous or early successive bilinguals who became German-dominant in childhood. 6 Due to various reasons (remigration, changes to the family constellations, professional reasons, among others), the degree of contact with either EP or German in daily life is diverse across speakers. The late L2 learners started to acquire German in a classroom setting after age 10. All late learners are highly proficient in German, either because they studied German at the university or moved to Germany or Switzerland for professional or personal reasons. At the moment of testing, 7 participants were living in Germany, 2 in a German Swiss canton, 12 in Portugal and one participant moved recently from Germany to Northern Spain (Galicia). 7 All participants signed informed consents for experimentation with human subjects previously approved by the Ethics Council of the University of Minho (CEICSH 120/2020) through Google Forms. Materials Measurement of language balance We assessed proficiency in each language through the DIALANG Vocabulary Size Placement Test (VSPT, version 1) for EP and German. Note that the lexical competence has been shown to be a reliable predictor of language proficiency (Bonvin et al., 2021; Laufer & Nation, 1999; Treffers-Daller & Korybski, 2016), since the learners’lexical knowledge grows when proficiency increases, and adequate lexical knowledge is a prerequisite of effective language use. The DIALANG VSPT is a questionnaire that assesses lexical competence through a list of 75 words, of which 50 are real words and 25 are pseudowords (for a more detailed explanation of the concept of language proficiency and balance as well as of supporting evidence of the DIALANG VSPT as a reliable proficiency indicator see Appendix S1). Participants were requested to indicate whether or not each word was an existing word in EP or German. Following Alderson (2005), the test score was computed based on the total of words correctly identified as either real words or pseudo-words. A measurement of language balance was obtained by calculating between-language subtractive differentials, i.e., the score obtained for German was subtracted from that of EP. Negative values indicate higher proficiency in EP and positive values in German (e.g., German: 62 - EP: 72; dif: -10). Values close to zero indicates high language balance. Complementing this, we asked participants to self-assess their proficiency in Portuguese and German, on a scale from 1 to 7 in speaking and writing. For the quantification we added the ratings for both skills and obtained a total self-assessment score on a scale from 1 to 14 for both languages. We then computed the differential between both language scores to define a value for selfassessed relative proficiency. Again, negative values indicate higher self-estimated proficiency in Portuguese and positive values in German (e.g., Portuguese speaking: 7 + Portuguese writing: 6 = 13; German speaking: 5 + German writing: 5 = 10, dif: - 3). A positive strong correlation between the language balance score obtained through the DIALANG VSPT and the selfassessment scores will further support the reliability of the DIALANG VSPT as a proficiency test (for more detail, check footnote 6 in the Results section; see also Flores et al., 2022). Stimuli We selected 180 EP inanimate nouns from the P-PAL database (for all the stimuli, see Appendix S2; Soares et al., 2018) and translated them into German. They were selected by taking into account the gender value in both languages, so that we had the same number of stimuli in each of these 6 translation types: heterogeneric feminine-masculine (“abóbora”[F] in EP, “Kürbis”in German [M], “pumpkin”), heterogeneric masculine-feminine (“journal”[M] in EP, “Zeitung”in German [F], “newspaper”), homogeneric feminine (“cenoura”in EP, “Karotte”in German, “carrot”), homogeneric masculine (“bosque”in Portuguese, “Wald”in German, “forest”), feminine-neuter (“perna”in EP, “Bein”in German, “leg”), and masculine-neuter (“carro”in EP, “Auto”in German, “car”). We avoided EP nouns with more than one German translation that had similar frequencies of use according to SUBTLEX-DE (Brysbaert et al., 2011). Besides, we avoided nouns in German that could also be verbs (e.g., “Leben”[life/to live]), nouns with high positive or negative affective valence related to death or sexuality (e.g., corpse, death, penis, etc.), German nouns that were cognates in English (e.g., “Butter”), and nouns that had more than one possible translation, when these diverged in gender within that language (e.g., “Miete” [rent, F] in German translates to “renda”[F] or “aluguer”[M] in EP). In terms of ortho-phonological gender transparency, we did not select any irregular nouns in EP, and included 144 transparent nouns and 36 opaque nouns evenly distributed across the six translation types (24 transparent nouns and 6 opaque nouns per type). We created two different blocks as a function of translation direction (EP to German and vice-versa), each composed of 90 stimuli, so that the presentation of both blocks was counterbalanced across participants. All EP nouns were controlled through a one-way ANOVAs across the 6 translation types for per million and logarithmic frequency, number of phonological and orthographic neighbours, number of letters, and mean logarithmic bigram frequency, taken from P-PAL (all ps > .247, Soares et al., 2018), logarithmic frequency, taken from SUBTLEX-PT (p= .544, Soares et al., 2015), and subjective frequency, concreteness, and imageability, taken from the Minho Word Pool (all ps> .525, Soares et al., 2017). See mean values in Table S1 of Supplementary Materials. German nouns were controlled across the six translation types through a one-way ANOVA for absolute logarithmic frequency, number of letters, logarithmic number of neighbours based on the Levenshtein distance, initial logarithmic bigram frequency (normalized), and familiarity, as taken from dlexDB (all ps> .091; Heister et al., 2011), and logarithmic frequency as taken 6 Research has shown that in early stages of language development simultaneous and early successive bilingual language acquisition may show developmental differences (e.g., Meisel, 2008); but these differences are overcome in older ages so that the language competence of simultaneous and early L2 speakers of a given language may become indistinguishable at least at adolescence (Montrul, 2016). Flores (2020), for instance, did not find any AoA effects in the competence of the Portuguese–German speakers analysed in her attrition study. Since the early bilinguals analysed in the present study are adults, who lived for an extended period of time in a German environment, there is no empirical support/evidence to further separate simultaneous from early acquirers of German. In fact, a clear separation between both acquisition types is typically not possible in heritage speakers who were born in the host country because it is hard (almost impossible) to determine the exact onset of exposure to the majority language of immigrant infants who are raised in a minority language environment (Montrul, 2016). 7 The location where the participants were recruited (i.e., their place of residence) was included in the final model, but it had no significant effect ( p> .05). The other effects in the model remained unchanged. We also compared the models that included this variable with those that did not and found that none of the comparisons were significant (all ps> .05). As a result, this variable was not included in the models presented in the manuscript. Bilingualism: Language and Cognition 7 https://doi.org/10.1017/S1366728923000378 Published online by Cambridge University Press from SUBTLEX-DE (Brysbaert et al., 2011;p= .29). See Table S2 of Supplementary Materials for means and standard errors. Translation pairs across translation types were controlled for equivalent measures –namely, number of letters, and subjective frequency/familiarity ( ps > .563). Although logarithmic frequency (SUBTLEX-PT and SUBTLEX-DE) showed significant differences, these differences were not between the conditions that subsequently showed significant results –namely, gender incongruency vs. congruency ( ps > .143). The translations were also controlled for orthographic overlap using the NIM database (Guasch et al., 2013) and phonological overlap using the PHOR-in-One database (Costa et al., 2021; all ps > .174). See Table S3 of Supplementary Materials for means and standard errors of overlap measures. Conditions regarding the stimuli were created taking into consideration the factors of Gender Congruency (gender congruent, gender incongruent), Target Gender (masculine, feminine), and Target Language (EP, German). Note that target gender and target language refer to the gender of the noun to be produced, and the language to be produced, respectively. Importantly, the factor of gender congruency included the four experimental conditions: heterogeneric masculine and feminine (incongruent) and homogeneric masculine and feminine (congruent) nouns. This allowed us to make a direct comparison between the gender systems of EP and German. Neuter gender in German would constitute a third category, in which rather than gender incongruency there is a situation in which one gender value does not exist in the other language, and this may change the representation of that value and the interaction between languages when it comes to its retrieval (for more information on the bilingual representation of differing gender nodes such as that of German and Spanish bilinguals in regard to the neuter node, see Klassen, 2016, and Klassen et al., 2022) . The scope of our study is to replicate the cross-linguistic GCE and test the activation of gender nodes and the processes of competition that may arise between them depending on language proficiency. These aims can be fulfilled with heterogeneric and homogeneric nouns, avoiding a greater degree of complexity in the experimental design, as we did. Nevertheless, neuter nouns were included among our stimuli in order to avoid artificial contexts that might in some way influence the performance of our participants. Procedure The experiment was conducted online due to the public international health emergency caused by the COVID-19 pandemic. We followed Burke and James’(2006) recommendations for online research and data collection. Participants started by filling out the LHQ (Li et al., 2020) using Google Forms. This questionnaire allowed us to explore their linguistic background and check their knowledge of any other gendered language. Afterwards, links were sent for the EP and German versions of the DIALANG, a standardized lexical test to objectively assess the proficiency in each of these languages while guaranteeing a high degree of validity and reliability (Alderson, 2005). The task was timed through the Google Add-on Quilgo, which, in addition to the timer, set for 7 minutes in total, allows for screen tracking, thus informing us if the participants kept focused on the task. Only one participant was excluded due to unfocussed participation. Participants then received a link that opened the experiment in a browser. The experiment was programmed using the JavaScript library jsPsych (de Leeuw, 2015). Two blocks of 90 nouns each (90 in EP and 90 in German) were created, but its order of presentation and the language to be produced (i.e., the target language) was counterbalanced, so that we had four different links depending on these two factors. After clicking on the link, the participant first read the instructions regarding the procedure for the whole experiment. These instructions appeared in the target language of the first sub-block. Then, a familiarization phase started in which the 90 translation pairs were sequentially presented, one by one. Participants controlled the presentation of the stimuli using the spacebar. The aim of the familiarization phase was to decrease mistranslations and non-responses. After that, participants tested their microphone, following instructions on the screen, and once checked, instructions appeared for the translation of the first block. They were asked to translate each noun into the target language as fast as possible, avoiding mistakes and trying to speak loudly and clearly into the microphone. Upon starting, participants went through a session of eight training items (different from the experimental items), then the experimental trials started. Each experimental trial had the following structure: a fixation point (+) at the centre of the computer screen, for 500 ms; the target noun for 3,000 ms or until response; a blank space for 500 ms as an inter-trial interval. Trials were presented randomly per participant. For the second sub-block, instructions appeared in the other language (which would be the new target language). The familiarization phase and translation task occurred again with the remaining 90 nouns. Responses were recorded and saved in a private directory on the University of Rovira i Virgili server. RTs were calculated offline from the presentation of the noun to be translated to the onset of the translation response using the PRAAT software (Boersma & Weenink, 2018). The session for each block lasted approximately 20 min. Results We removed the RTs of incorrect responses (18.04% data points), RTs above 3,000 ms and RTs that exceeded 2.5 SD of each participant’s mean (2.35% data points). We also removed the data from ten participants that made more than 40% of incorrect responses. Hence, the final sample was composed of 64 participants. Table 1 gives an overview over the predictor variables for these 64 participants –namely, AoA, self-rating in each language, self-assessed relative proficiency, proficiency in each language and language balance score. RTs were analysed using linear mixed-effect models (e.g., Baayen, 2008; Baayen et al., 2008). To this end, we used the lme4 package of R (Bates et al., 2015). We created a fixed structure model to examine the hypotheses of the study, with the inverse of RTs (-1000/RT) as the dependent variable. As fixed effects, the model included the triple interaction and second-order interactions between Gender Congruency (GC, GI), Target Language (language that was produced: Portuguese or German) and absolute Language Balance Score (through the DIALANG) 8 , the triple interaction and second-order interactions between Gender Congruency, absolute Language Balance Score and Age of Acquisition (AoA) of the German Language, the interaction 8 It should be noted that the self-assessed relative proficiency differential and the DIALANG differential were highly correlated, r= .79, p< .001. We therefore decided to introduce the DIALANG differential in the models, instead of the self-assessed, as it is a more objective measure of the participants’language proficiency. 8 Ana Rita Sá‐Leite et al. https://doi.org/10.1017/S1366728923000378 Published online by Cambridge University Press between Block Order (first and second) and Target Language, and, finally, Target Gender (gender of the noun that was produced: Feminine or Masculine). Continuous variables were centered and transformed into Z-scores. In addition, following the guidelines of Schad et al. (2020), all dichotomous variables were coded using sum contrast coding (-0.5 for the first level and +0.5 for the second level of each factor); Gender Congruency: GC (-0.5), GI (+0.5); Target Language: GER (-0.5), PT (+0.5); Target Gender: FEM (-0.5), MASC (+0.5), and Block Order: first (-0.5), second (+0.5). We also examined the multicollinearity of the fixed effects introduced in the model (R VIF function). All VIF values were less than 3, suggesting non multicollinearity (Zuur et al., 2010). Participants and words were included as grouping factors for random effects. We followed a maximal random-effects structure (Barr et al., 2013) by adding as random slopes the most complex structure that allowed convergence. We incorporated Target Language and Gender Congruency into the random slope of participants, and Target Gender into the random slope of words. The structure of the models for evaluating the RTs in Portuguese and German naming was the same as above, but excluding Target Language for participants. The significance of interactions was determined using log-likelihood ratio tests (R ANOVA function). We assessed the contribution of each interaction by comparing a model that included them with another model in which they were not included. We also report the results of the t-test analyses for the coefficient estimates of fixed effects and interactions. To this end, we used Satterthwaite’s approximations to the degrees of Table 1. Sociolinguistic background of the 64 analyzed participants and DIALANG results AoA of seven or earlier (n = 42) AoA higher than ten (n = 22) MSD Range MSD Range min max min max Age 38.45 9.54 18 55 38.27 10.22 21 63 Average AoA 2.1 2.01 0 7 15.77 2.67 10 25 Self-rating Portuguese 12.55 1.82 6 14 13.55 1.14 9 14 Self-rating German 12.48 1.86 6 14 10.55 2.39 6 14 Self-assessed relative proficiency −0.07 1.63 −53−3 2.64 −83 Average European Portuguese proficiency (Dialang) 64.1 4.7 51 72 67.68 3.23 62 75 Average German proficiency (Dialang) 68 5.04 52 75 61.36 7.46 49 71 Language balance score (Dialang) 3.88 6.98 −17 20 −6.32 7.73 −26 4 Note. M = mean; SD = standard deviation; min = minimum; max = maximum. Negative values indicate higher proficiency in Portuguese in Self-assessed relative proficiency and Language Balance Score (DIALANG). Table 2. Results of the linear mixed-effects model Predictors Estimate SE t p Intercept −0.93 0.01 −64.86 <.001 Gender Congruency 0.01 0.02 0.40 0.691 Target Gender 0.01 0.01 0.90 0.371 Target Language −0.02 0.01 −3.78 <.001 Language Balance Score 0.01 0.01 0.64 0.523 AoA German −0.00 0.01 −0.23 0.815 AoA German * Gender Congruency −0.01 0.00 −1.13 0.259 AoA German * Language Balance Score −0.01 0.01 −0.73 0.464 AoA German * Language Balance Score * Gender Congruency 0.00 0.01 0.33 0.743 Gender Congruency * Target Language −0.01 0.01 −0.57 0.567 Gender Congruency * Language Balance Score 0.01 0.01 1.24 0.217 Language Balance Score * Target Language −0.01 0.01 −0.94 0.345 Gender Congruency * Language Balance Score * Target Language 0.02 0.01 2.12 0.034 Table 3. Mean RTs and standard errors Target Language Target Gender Gender Congruency Mean RTs SE GER F GC 1121 10.5 GER F GI 1130 11.1 GER M GC 1161 11.4 GER M GI 1152 10.6 PT F GC 1092 10.1 PT F GI 1124 10.5 PT M GC 1138 10.8 PT M GI 1114 11.3 Note. Results reported for the conditions of Gender Congruency taken into consideration Target Language and Target Gender. Bilingualism: Language and Cognition 9 https://doi.org/10.1017/S1366728923000378 Published online by Cambridge University Press