Within-language lexical interference can be resolved in a similar way to between-language interference

Iva Ivanova; Dacia Carolina Hernandez

doi:10.1016/j.cognition.2021.104760

. Author manuscript; available in PMC: 2021 Sep 1.

Published in final edited form as: Cognition. 2021 Jul 1;214:104760. doi: 10.1016/j.cognition.2021.104760

Within-language lexical interference can be resolved in a similar way to between-language interference

Iva Ivanova ¹, Dacia Carolina Hernandez ¹

PMCID: PMC8335802 NIHMSID: NIHMS1704182 PMID: 34218002

Abstract

This study asks if monolinguals can resolve lexical interference within a language with mechanisms similar to those used by bilinguals to resolve interference across languages. These mechanisms are known as bilingual language control, are assumed to be at least in part top-down, and are typically studied with cued language mixing, a version of which we use here. Balanced (Experiment 1) and nonbalanced Spanish-English bilinguals (Experiment 2) named pictures in each of their languages. English monolinguals from two different American cities (Experiments 3 and 4) named pictures in English only with either basic-level (e.g., shoe) or subordinate names (e.g., sneaker). All experiments were identically structured and began with blocked naming in each language or name type, followed by trial-level switching between the two languages or name types, followed again by blocked naming. We analysed switching, mixing and (introduced here) post-mixing costs, dominance effects and repetition benefits. In the bilingual experiments, we found some signs of dominant deprioritization, the behavioral hallmark of bilingual language control: larger costs for dominant- than for nondominant-language names. Crucially, in the monolingual experiments, we also found signs of dominant deprioritization: larger costs for basic-level than for subordinate names. Unexpectedly and only in the monolingual experiments, we also found a complete dominance reversal: Basic-level names (which otherwise behaved as dominant) were produced more slowly overall than subordinate names. Taken together, these results are hard to explain with the bottom-up mechanisms typically assumed for monolingual interference resolution. We thus conclude that top-down mechanisms might (sometimes) be involved in lexical interference resolution not only between languages but also within a language.

Keywords: language production, bilingual language control, picture naming, mixing costs, switching costs, dominance effect

It is near impossible to hear a Spanish-English bilingual say “mesa” when they intend to say “table”: Bilinguals possess efficient and well-attested language control mechanisms to avoid saying words in the wrong language when they may not be understood (Gollan, Sandoval, & Salmon, 2011; Poulisse, 1999). But control of word selection seems no less necessary when conversing in a single language: Saying “dog” instead of “poodle” might end up ambiguous, “chair” instead of “table” puzzling, and “half-assed” instead of “inadequate” plain unwise in view of one’s chances of getting most jobs. Can monolinguals use within a language the same lexical interference resolution mechanisms that bilinguals use across languages? That is, are there language- and speaker-general mechanisms of control over word selection that can be applied both across and within languages? We address these questions here by tracing whether established behavioral signatures of language control in bilinguals generalize to monolinguals naming pictures with basic-level (e.g., shoe) and subordinate names (e.g., sneaker) in their only language.

In asking these questions, we acknowledge that there are many differences between the language and cognitive systems of individuals who typically use two languages to communicate and those who typically use one. There is a lot of evidence that bilingualism brings cognitive and neural adaptations (Abutalebi & Green, 2007; Bialystok, Craik, & Freedman, 2007; Bialystok, Craik, & Luk, 2008; Bialystok & Martin, 2004; Costa, Hernandez, & Sebastián-Gallés, 2008; but see Von Bastian, Souza, & Gade, 2016; Kousaie & Phillips, 2012; Paap & Greenberg, 2013). Further, two languages differ from each other in phonology and syntax (a difference absent within a single language), and they impose the need to “negotiate” which language is spoken when. The idea behind this study was not to underplay any of these differences, or the many benefits of being able to communicate in two languages. We (similarly to Dylman & Barry, 2018) simply aimed to see if there could be a certain commonality of mechanism applied to lexical representations within and between languages. Even if that were the case, there would still be an important difference between monolinguals and bilinguals: Such a process would be applied a lot more frequently by bilinguals (to prevent both between- and within-language interference) than by monolinguals (who only need to prevent within-language interference).

For the sake of clarity, we talk about dominant and nondominant languages instead of first and second languages (or L1 and L2) because the dominant language is typically different from the first-learned language for the population we study here. For consistency, we also use these terms to refer to L1 and L2 as used by other authors (Christoffels et al., 2016, Dylman & Barry, 2018, and Kroll & Stewart, 1994).

Resolving cross-language lexical interference in bilinguals

Preventing interference from the words belonging to the non-target language in bilingual speech is necessary under the assumptions that bilinguals’ two languages have a common conceptual system (for a review, see Francis, 2005), and shared concepts activate both a target word and its non-target translation equivalent (Colomé, 2001; Costa, Caramazza, & Sebastián-Gallés, 2000; Hoshino & Kroll, 2008). Most accounts assume that translation equivalents then compete for selection with target words (following e.g., Roelofs, 2003). This makes it non-trivial to explain how bilinguals manage to restrict production to a target language when required by the communicative situation (and produce extremely few wrong-language intrusions: Gollan et al., 2011; Poulisse, 1999).

Cross-language interference prevention in bilinguals is known as language control. Bilingual language control may function in several different ways. In the most established view, bilinguals avoid wrong-language intrusions by inhibiting the non-response language (Inhibitory Control Model: Green, 1998). A Supervisory Attentional System (as in Norman & Shallice, 1986) acts on language task schemas (the mental structure of different language activities such as “Speak Language X”) that in turn send inhibition to nontarget-language schemas (global inhibition) and nontarget-language lexical representations (local inhibition). Inhibition is proportional to the strength of the language it acts on: A more dominant language (more likely to cause interference) is inhibited more strongly than a non-dominant language (less likely to cause interference). In non-inhibitory accounts, language control happens through context-dependent adjustments of the activation levels (Finkbeiner, Gollan, & Caramazza, 2006; La Heij, 2005; Poulisse & Bongaerts, 1994) or lexical selection thresholds (Branzi, Martin, Abutalebi, & Costa, 2014) of words in either or both the target and non-target languages in situations of dual language activation (see also Grosjean, 2001; Philipp, Gade, & Koch, 2007; Runnqvist, Strijkers, & Costa, 2019). Non-inhibitory language control can also be achieved if translation equivalents do not compete for selection (Costa, Caramazza, & Sebastián-Gallés, 1999), a state possibly attainable only with high proficiency in at least two languages (Costa, Santesteban, & Ivanova, 2006). Regardless of the precise mechanism, bilingual language control is usually assumed to involve a top-down component (Green, 1998; for neural evidence, see Abutalebi, 2008; Abutalebi, Annoni, Zimine, Pegna, Seghier et al.; 2008; Blanco-Elorrieta & Pylkkänen, 2016; Seo, Stocco, & Prat, 2018; overview in Abutalebi & Green, 2007; Calabria, Costa, Green, & Abutalebi, 2018). One exception is the account proposed by Runnqvist, Strijkers, Alario, and Costa (2012), which we discuss below.

Indicators of bilingual language control

The most consistent behavioral manifestation of bilingual language control is a deprioritization of the dominant language for nonbalanced bilinguals in a dual-language context, relative to when the dominant language is produced without prior (recent) activation of the nondominant language. Dominant deprioritization amounts to slower responses in timed picture naming (Branzi et al., 2014; Costa & Santesteban, 2004; Macizo, Bajo, & Paolieri, 2012; Meuter & Allport, 1999; Philipp et al., 2007; Verhoef, Roelofs, & Chwilla, 2009), reduced lexical accessibility in verbal fluency tasks (Van Assche, Duyck, & Gollan, 2013) and increased number of intrusion errors in paragraph reading (Gollan & Goldrick, 2016; Gollan, Schotter, Gomez, Murillo, & Rayner, 2014). We use the term “deprioritization” in a theory-neutral sense that includes the possibility for a “boost” of the nondominant language when it is target instead of a “hit” to the dominant language when it is non-target (e.g., Philipp et al., 2007; Runnqvist, Strijkers, & Costa, 2019).

Some of the most frequent and robust demonstrations of dominant deprioritization come from cued picture naming in two languages. Such tasks have produced three relatively consistent patterns of dominant deprioritization, each of which is considered in the literature a behavioral “signature” of language control. The first is switching costs. They are measured in a mixed-language number- or picture-naming task (e.g., Meuter & Allport, 1999; Costa & Santesteban, 2004; review in Declerck & Philipp, 2015). One version of this task involves (predictably or unpredictably) switching languages every few trials. Naming a picture in a language different from that of the preceding trials (i.e., switching) results in a reaction time cost known as switching cost¹, assumed to result from the time needed to disengage from the preceding task and recuperate the current target language from a previously applied control mechanism. In the majority of studies, switching costs were larger for the dominant language of nonbalanced bilinguals, i.e., showed dominant deprioritization (Costa & Santesteban, 2004; Jin, Zhang, & Li, 2014; Macizo et al., 2012; Meuter & Allport, 1999; Peeters, Runnqvist, Bertrand, & Grainger, 2014; Philipp et al., 2007; Verhoef et al., 2009; review in Bobb & Wodniecka, 2013; although some studies have found symmetrical switching costs for two languages of unequal dominance, Christoffels, Firk, & Schiller, 2007; Verhoef, Roelofs, & Chwilla, 2010, or larger switching costs for the non-dominant language, Declerck, Stephan, Koch, & Phillip, 2015). Note that symmetrical switching costs for the two languages of balanced bilinguals (Costa & Santesteban, 2004; Costa et al., 2006) are predicted because two similarly dominant languages would necessitate language control mechanisms of similar strength.

Adding single-language blocked naming for the two languages to the switching task described above also allows to measure mixing costs. Mixing costs are determined by subtracting naming latencies during blocked naming from naming latencies on non-switch trials in a mixed-language block (i.e., naming a picture or digit in a given language after having named the previous picture or digit in the same language). Mixing costs are thought to reflect the global costs of maintaining readiness to produce two different name types in mixed blocks, and are considered another indicator of language control processes. Typically, mixing costs also show an asymmetry for non-balanced bilinguals, such that there are larger mixing costs in the dominant than in the non-dominant language(s) (Christoffels et al., 2007; Gollan & Ferreira, 2009; de Bruin, Roelofs, Dijkstra, & FitzPatrick, 2014; Mosca & Clahsen, 2016; Peeters & Dijkstra, 2018). Of note, switching costs and mixing costs appear to be complementary, such that only one marker shows dominant deprioritization in tasks where both are measured (mixing costs showed dominant deprioritization: Christoffels et al., 2007; de Bruin et al., 2014; Gollan & Ferreira, 2009; Mosca & Clahsen, 2016; Peeters, & Dijkstra, 2018; switching costs showed dominant deprioritization: Declerck, Philipp, & Koch, 2013; Wang, Kuhl, Chen, & Dong, 2009). In balanced bilinguals, mixing costs should be symmetrical for the same reason as switching costs: Languages of similar dominance would need the application of similar amounts of control.

A third indicator of control processes during cued mixed-language picture naming is the language dominance effect, which refers to the overall speed difference between the dominant and the non-dominant language. Typically and expectedly, dominant naming is faster when it precedes non-dominant naming (Christoffels et al., 2007; for reviews, see Hanulová, Davidson, & Indefrey, 2011; Runnqvist, Strijkers, Sadat, & Costa, 2011), because dominant-language words are more robustly represented and are assumed to have higher resting activation or lower selection thresholds. However, in dual-language situations, the overall naming speeds for the two languages may become similar (Calabria, Hernández, Branzi, & Costa, 2012; Declerck et al., 2013, Experiment 3; Filippi, Karaminis, & Thomas, 2014; Fink & Goldrick, 2015; Kleinman & Gollan, 2016), or the non-dominant language may even be produced more quickly overall (a reversed dominance effect: Costa & Santesteban, 2004; Costa et al., 2006; Christoffels et al., 2007; Gollan & Ferreira, 2009; Li & Gollan, 2018; Declerck, Kleinman, & Gollan, 2020; Peeters & Dijkstra, 2018; Verhoef et al., 2009; 2010; trend in Mosca & Clahsen, 2016). Equalized or reversed dominance has been interpreted as global deprioritization of the dominant language. We note that the three language control indicators discussed so far are not independent of one another, and the three may pattern together differently across studies because of differences in tasks (e.g., length, number of stimuli, number of repetitions, switch-no switch ratio, stimulus valence) and participants’ language history.

Finally, we reasoned that a task with single-language blocked naming both before and after trial-level switching (see Christoffels, Ganushchak, & La Heij, 2016; Gollan, Kleinman, & Wierenga, 2015; Prior & Gollan, 2011) would allow measuring another type of costs. We thus introduce post-mixing costs: blocked naming latencies after trial-level switching minus blocked naming latencies before trial-level switching. These costs could be another marker of dominant deprioritization that reflects sustained effects of language mixing independent of the need for task-set reconfiguration.² They also have the advantage of comparing the same types of trial, blocked naming.

Resolving within-language lexical interference in monolinguals

The problem of shared concepts highly activating two lexical items is not limited to bilinguals. It also occurs for synonyms (albeit to a lesser extent because not all words have close synonyms), and, at least partially, to words that share aspects of meaning such as register equivalents (e.g., dog – pooch) or words of different levels of specificity (e.g., dog – Dalmatian). However, the bottom-up mechanisms proposed for monolingual lexical interference resolution (reviewed below) are qualitatively different from the top-down mechanisms typically assumed for bilinguals (e.g., Abutalebi et al., 2008; Abutalebi & Green, 2007; Blanco-Elorrieta & Pylkkänen, 2016; Calabria et al., 2018; Green, 1998; Seo et al., 2018).

Within-language (monolingual) lexical interference resolution is explained by two broad types of account, with or without competition for selection. On the traditional account, words within a language compete for selection (and greater competition results in slower and more effortful production: e.g., Starreveld & La Heij, 1995; Wheeldon & Monsell, 1994). One proposed solution to keeping in check competitors’ activation levels is lateral inhibition (inhibition originating from a particular representation in proportion to its own current activation, and not from an external system, such as in the bilingual Inhibitory Control Model of Green, 1998). Activated lemmas (abstract word representations) send lateral inhibition to other activated lemmas to prevent interference (Cutting & Ferreira, 1999; Howard, Nickels, Coltheart, & Cole-Virtue, 2006). On an alternative account, there is no lateral inhibition, and instead the system examines all activated words and selects the one whose activation exceeds the activation of all others by some critical amount (Levelt et al., 1999; Roelofs, 1992). A further recent hypothesis is that the activation differences between target and non-target words that would produce competition result from a flexible criterion, adjusted in response to factors such as task goals (Nozari & Hepner, 2018). In the way we understand this account, goals would be evaluated, and the criterion set, by top-down decision-making processes – but competition itself would be resolved by the language-internal bottom-up processes described above.

Without competition, words’ ease and speed of selection may hinge on the speed with which target and non-target words reach the (bottleneck) articulation stage, and with which non-target words can be rejected (Response-exclusion Account, Mahon, Costa, Peterson, Vargas, & Caramazza, 2007). There is also no competition for selection in the incremental learning model of Oppenheim et al. (2010), in which a naming instance (e.g., horse) strengthens the conceptual-lexical connections of that representation, and weakens the connections of all related representations (e.g., cow) in proportion to their relatedness. Of relevance here, no-competition accounts also involve mechanisms that are qualitatively different from the ones proposed for resolving between-language interference (e.g., Abutalebi & Green, 2007; Green, 1998).

Speaker- and language-general mechanisms for interference resolution?

But is it possible that monolinguals and bilinguals use the same control mechanisms within- and across languages? The idea of shared features of the monolingual and bilingual lexical selection systems has been put forward by several authors, but the evidence is mixed. On the one hand, Dylman and Barry (2018) showed evidence for commonalities between bilingual and monolingual lexical selection. Three groups of bilinguals (Spanish-English, Swedish-English and English-French) and two groups of British English monolinguals named pictures while ignoring superimposed distractor words. For bilinguals, there were targets and distractors in both their first and dominant, and in their second and weaker language. For monolinguals, targets and distractors were either common names (e.g., dog) or alternative, less frequent co-ordinate synonyms (e.g., hound). There was facilitation from both translation-equivalent and synonym distractors, but more weakly represented names (nondominant-language names for bilinguals and alternative synonyms for monolinguals) benefitted more from it than more robustly represented names (dominant-language names for bilinguals, and common synonyms for monolinguals). While acknowledging that translation equivalents and synonyms differ in some respects, the authors concluded that they are represented similarly at some level. These findings suggest that there can be similarities between single and dual language systems.³

On the other hand, Melinger (2018) showed that distractor words belonging to a different dialect (or a different register) produced interference, in contrast to the (translation-)facilitation effects in Dylman & Barry (2018) and Costa et al. (1999). Melinger (2018) thus concluded that dialects are represented differently from languages and draw on different control mechanisms, specific to within-language control. The opposing within-language distractor effects (competition from category-coordinate different dialect distractors and no competition from synonyms) were reconciled by Melinger (2021), who replicated them in the same experimental paradigm. She proposed that the effects depend on the number of currently activated strong lexical competitors, which create a trade-off between lexical competition (when there are many lexical competitors) and conceptual facilitation (when there is only one lexical competitor). Important here, this proposal situates both types of effect within the same system, and attributes the different polarity of effects to representational differences.

Other studies have also concluded that single- and dual-language control use (at least in part) similar mechanisms. Declerck, Grainger, Koch, & Philipp (2017; Experiment 3) compared the switching costs between language and within language (via switching between picture naming and category naming). The two types of switching costs were correlated and the costs in naming latencies did not differ between the two tasks (although the costs in error rates did). Further, Kirk, Kempe, Scott-Brown, Philipp, and Declerck (2018) showed similar switch-cost patterns for dialect switching as for language switching. In contrast to Melinger (2018), they concluded that monolingual dialect speakers recruit control mechanisms in similar ways as bilingual speakers.

Further evidence for shared mechanisms of control of single and dual-language lexical selection was provided by Finkbeiner, Almeida, Janssen, and Caramazza (2006), who found dominant deprioritization in a single-language context by monolinguals. In this study, monolinguals reading “faster” words (e.g., dog, house, stone) incurred larger switching costs than reading “slower” words (e.g., puppy, cottage, pebble) during switching between word reading and color naming (which the authors interpreted as evidence against bilingual inhibitory control).

Evidence for single- and dual-language control mechanisms that are in part similar and in part different was shown by Declerck, Ivanova, Duñabeitia, and Grainger (2020). These authors compared the trial-level costs incurred by bilinguals when switching between their two languages (French and English) and the costs incurred by these same bilinguals when switching between formal and informal words (e.g., dog versus pooch) in their dominant language (French). The switching costs were comparable, symmetrical in both cases, and correlated. However, increasing the cue-to-stimulus interval reduced the between-language switching costs (a reduction the authors attributed to the opportunity for phonological preparation of the switch to the other language) but left unaffected the within-language switching costs (where there is no opportunity to benefit from the preparation of a different phonological system).

On the other hand, neural differences between processing in a dual- and a single-language context were shown by Abutalebi, Annoni, Zimine, Pegna, Seghier et al. (2008). Switching between naming objects in two languages caused increased activation in the left caudate and anterior cingulate cortex, relative to switching between naming objects and their associated verbs in a single language by German-French bilinguals. However, the need to prevent interference between object names and verbs (which are not used in the same grammatical context and thus might never compete for selection in real-life language use) might not be as high as between object names in two languages (which mean the same thing and fit within the same grammatical context).⁴

An insightful investigation of the nature of the possibly shared mechanisms of between- and within-language control was conducted by Runnqvist et al. (2012). These authors tested models of bilingual language control – the Inhibitory Control Model (Green, 1998) and language-specific lexical selection (Costa et al., 1999) – with the cumulative semantic interference (CSI) paradigm. They found that naming latencies increased with the naming of each subsequent category member – a cumulative semantic interference effect that was equivalent when the response language alternated within a run and when it was kept constant. These results are inconsistent with either of the bilingual control models because they predict that the interference effect would be reduced in half in alternating runs (because inhibiting or not attending to non-response language representations would make them non-competitors). Instead, the authors argued that their results are most parsimoniously explained with the bottom up mechanisms proposed for monolingual lexical selection by Howard et al. (2006) and Oppenheim et al., 2010 (see also Runnqvist et al., 2019, for relevant discussion)⁵. This proposal successfully accounts for target selection within and across languages and is appealing in its parsimony.

However, we note that this proposal does not, without additional assumptions, seem to explain how dominant deprioritization occurs. This is because the strength of bottom-up mechanisms such as activation and lateral inhibition is assumed to depend on a word’s own activation levels. Nondominant-language representations would have lower activation levels than dominant language words. Consequently, they would not be very strong competitors (and neither could they win a competition solely based on activation levels). They would also not be able to spread stronger lateral inhibition than the one coming from dominant-language representations (such that could disproportionally disadvantage dominant-language words). Going back to our review above, the only bottom-up mechanism able to explain dominant deprioritization in certain situations is the Response-exclusion Account of Mahon et al. (2007) – the assumption that when dominant representations reach the articulatory buffer too fast in a situation of potential conflict, they are subjected to enhanced scrutiny, disproportionally slowing them down. Still, this account also appeared unable to explain the totality of results from the current study; we explain why in the General Discussion.

The Present Study

We ask if monolinguals can use the same mechanism to prevent within-language interference from highly activated semantically-related lexical items as bilinguals use to prevent between-language interference from highly activated translation equivalents. The existing evidence on the issue is mixed. Here, we analyze the language control indicators of monolinguals performing a within-language switching task. Of main interest is if monolinguals will show signs of dominant deprioritization, a robust consequence of language control in bilinguals. To have a bilingual benchmark in the current context, we also analyzed the language control indicators of bilinguals who performed a between-language switching task that was designed to be maximally comparable to the within-language task. Balanced (Experiment 1) and non-balanced bilinguals (Experiment 2) named colored photographs in English and Spanish, and monolinguals from two different American cities (Experiments 3 and 4) named photographs with basic-level (e.g., shoe) and subordinate names (e.g., sneaker) in English. The monolinguals in Experiment 4 were from the same city as the bilinguals. The between- and within-language tasks were identically structured and had a sandwich design. Each task began with blocked naming in each language or name type (pre-mix blocks), followed by a mixed block (trial-level switching between languages or name types), followed again by blocked naming in each language or name type (post-mix blocks).

For each experiment, the measure of interest is naming latencies, and we analyze the previously discussed indicators of language control: switching and mixing costs (to which we add an exploration of post-mixing costs), and the language or name-type dominance effect. In nonbalanced bilinguals, all of these markers index the presence of language control mechanisms through dominant deprioritization, and at least one of them should show dominant deprioritization when all three are measured. For balanced bilinguals, on the other hand, dominant deprioritization effects should be absent (or, to the extent that bilinguals are rarely perfectly balanced, at least detectably weaker than for nonbalanced bilinguals). Thus, based on established effects in prior literature, we predict significant and symmetrical switching and mixing costs for balanced bilinguals and asymmetrical switching or mixing costs for nonbalanced bilinguals.

The empirical question of main interest in this study is how control markers would pattern for monolinguals. Following the logic of Dylman and Barry (2018), we predict that monolinguals would show dominant deprioritization (of the more dominant basic-level names) if they employ similar control mechanisms to bilinguals in the current context.

We consider that basic-level names are dominant because they tend to be more frequent and familiar, and subordinate names are nondominant because they tend to be less frequent and familiar, but we also analyze repetition benefits of each name to obtain further evidence of name dominance. Repetition benefits and repetition priming are typically attributed to both short- and long-term strengthening of connections between representations in the language system (e.g., between semantic and lexical, or lexical and phonological representations; e.g., Barry, Hirsh, Johnston, & Williams, 2001). Of interest here, weaker representations – both within language and across languages – benefit more from such strengthening than stronger representations (low-frequency names: Griffin & Bock, 1998; nondominant language names: Kleinman & Gollan, 2018; Misra, Guo, Bobb, & Kroll, 2012). The larger effect for weaker representations may be due to more room for improvement for representations further away from lexical selection thresholds, or larger adjustments of system weights (as in error-driven learning models, e.g., Oppenheim, Dell, & Schwartz, 2010) when a repetition amounts to a larger percentage of the total times a word is used. We also assumed that basic-level and subordinate names compete for selection (since they interfere with each other in picture-word tasks, Hantsch, Jescheniak, & Schriefers, 2005; Kuipers, La Heij, & Costa, 2006).

After conducting separate analyses of each experiment, we also conducted exploratory statistical comparisons between experiments (reported in brief in the main text and in full in Appendix C). We note, however, that our main predictions (presence or absence of dominant deprioritization for monolinguals) hinge on the analyses of the individual experiments. The comparison of main interest was between monolinguals and nonbalanced (rather than balanced) bilinguals because basic-level and subordinate names within a language should involve a similar (although perhaps less pronounced) representational imbalance to that between a dominant and a nondominant language. Balanced bilinguals were an important part of the study design as a demonstration that the current paradigm does not produce the same dominant deprioritization effects across the board but instead is sensitive to the theoretically predicted grading of effects for non-balanced and balanced bilinguals.

The present study adds to the investigation of the generality of language control systems in the following ways. First, it is the only study known to us to focus on dominant deprioritization within and across languages, a phenomenon not straightforwardly explained by the bottom-up models of monolingual lexical selection. Second, it provides a systematic examination of dominant deprioritization by analyzing four different markers of it (differently from previous studies of this type such as Declerck et al., 2017, 2020 and Kirk et al., 2018, who looked only at switching costs). Third, it looks at within-language switching between two different name types without the presence of another task (differently from some previous studies of this type which involved other tasks: category naming in Declerck et al., 2017, color naming in Finkbeiner et al., 2006). Fourth, it compares bilinguals with monolinguals (and not bilinguals within themselves) because bilinguals may have transferred their between-language control mechanisms to within-language control (unlike Declerck et al., 2017, 2020a, who compared single- and dual-language performance for bilinguals only).

Experiment 1: Language mixing in balanced bilinguals

Method

Participants.

Fifty-six balanced Spanish-English bilinguals, undergraduates at the University of Texas at El Paso (UTEP), participated for course credit or pay ($10/hour). The originally planned sample size, 48 participants per experiment, was based on the maximum in the range of sample sizes in published experiments of this type (N = 12–48; Costa & Santesteban, 2004; Declerck, Koch, & Philipp, 2012, Li & Gollan, 2018; Kirk et al., 2018; Peeters et al., 2014; Mosca & Clahsen, 2016; Philipp et al., 2007; Verhoef et al., 2009).⁶ The final sample of 56 participants (48 + 8) was obtained after reinclusion of eight balanced bilinguals whose data was originally excluded due to an oversight.

Bilinguals’ classification as balanced (Experiment 1) or nonbalanced (Experiment 2) was based on their English and Spanish scores on the Multilingual Naming Test (MINT), an objective productive vocabulary measure with 68 picture names of progressive difficulty (Gollan, Weissberger, Runnqvist, Montoya, & Cera, 2012). To be considered balanced, bilinguals had to be able to name not more than nine pictures in one language than in the other. On average, these bilinguals named three more pictures in English than in Spanish (range for English MINT scores minus Spanish MINT scores: −9 – 9). Participants’ language history in all experiments is reported in Table 1.

Table 1.

Language history characteristics of participants in all experiments.

	Balanced bilinguals, Exp. 1	Nonbalanced bilinguals, Exp. 2	Monolinguals (San Diego), Exp. 3	Monolinguals (El Paso), Exp. 4
Age of first exposure in years
English	6.1 (3.6)	^a3.1 (3.0)	0.15 (0.74)	1.0 (2.2)
Spanish	1.0 (1.3)	^a2.9 (3.9)	not collected separately
Other	14.4 (4.1), N = 30	13.3 (4.4), N = 23	^b11.4 (5.4), N = 46	^b11.5 (6.4), N = 15
% daily use now
English	62% (19%)	^a82% (17%)	95% (20%)	95% (17%)
Spanish	42% (21%)	^a23% (20%)	not collected separately
Other	12% (9%), N = 7	5% (5%), N = 3	^b,c8% (12%), N = 4	^b2% (3%), N = 4
% daily use as a child
English	35% (23%)	^a70% (27%)	95% (20%)	95% (20%)
Spanish	66% (22%)	^a38% (29%)	not collected separately
Other	none reported	10%, N = 1	^b6% (4%), N = 3	^b1% (3%) N = 4
Self-rated proficiency
English	9.0 (1.0)	^a9.7 (0.6)	10.0 (0.03)	9.8 (1.3)
Spanish	8.5 (1.4)	^a5.2 (2.2)	3.2 (1.7), N = 38	2.2 (1.7), N = 37
Other	3.2 (2.1), N = 26	2.8 (2.0), N = 19	2.3 (2.1), N = 22	2.5 (2.8), N = 23
Productive vocabulary (MINT, of 68)
English	57.8 (4.8)	62.1 (2.8)	Not collected	62.9 (3.1)
Spanish	55.2 (4.5)	29.6 (10.5)	Not collected	6.1 (7.7)

Open in a new tab

Note: Standard deviations are provided in parentheses.

N = 62 (the language history questionnaires of seven participants could not be uniquely identified).

Includes Spanish.

One participant reported speaking a language other than English 25% of the time.

Across Experiments 1 and 2 combined, 41 additional bilinguals were excluded from analyses. For most of them (N = 29), this was because they did not fit into the dominance group of either experiment. Twenty-two bilinguals were moderately English-dominant, naming on average 12 more pictures in English than in Spanish (range 10–15; the criterion for inclusion in Experiment 2 was naming more than 15 pictures in English than in Spanish). Part or all of the naming latency data for two of these participants were lost. Seven bilinguals were Spanish-dominant, naming on average 14 fewer pictures in English than in Spanish (range −10 – −23). In addition, twelve bilinguals were excluded because of missing latency data for the whole (N = 4) or more than half of the experiment (N = 5), missing MINT scores (N = 1), or because it was impossible to uniquely match their latency data, error rates and MINT (N = 2).

Materials.

The materials consisted of 12 colored photographs of common objects or animals inside a red or a blue frame to cue the target language (see Figure 1 for examples). The English names of the photographs (bag, bird, building, cheese, dress, fish, glass, horse, knife, painting, shoe, tree) and their Spanish translations (bolsa, pájaro, edificio, queso, vestido, pez, vaso, caballo, cuchillo, cuadro, zapato, árbol) were all non-cognates and of medium to high frequency (English: M = 64.48, SD = 26.17; Spanish: M = 47.62, SD = 26.90, t = 3.54, p = .005). Frequency-per-million values were obtained from the movie-subtitles corpora SUBTLEX-US for American English (Brysbaert & New, 2009; http://expsy.ugent.be/subtlexus/), and SUBTLEX-ESP for Spanish (Cuetos, Glez-Nosti, Barbón, & Brysbaert, 2011; http://crr.ugent.be/archives/679). (Note that the SUBTLEX-ESP database largely reflects Castilian Spanish use and is thus only our best approximation for the Mexican-Spanish-speaking population tested here.) The English names were shorter than the Spanish names, reflecting language-wide differences in word length (English: M = 4.08, SD = 1.73; Spanish: M = 5.50, SD = 1.38; length in phonemes, t = 4.71, p = .001). The higher frequency and shorter length of the English versus the Spanish names (a limitation of these stimuli) were expected to enlarge the inherent proficiency differences between the two languages (for nonbalanced bilinguals). However, we assume that lexical characteristics and proficiency have a common origin and influence lexical selection in the same way (as proposed by the Weaker Links account, Gollan, Montoya, Fennema-Notestine, & Morris, 2005; see also Strijkers, Costa, & Thierry, 2010, about a common origin of frequency and cognate effects). That is, we assume that the higher frequency and shorter length of the English relative to the Spanish names added quantitative but not qualitative processing differences between the names in the two languages.

Figure 1. — Examples of experimental pictures in all experiments. Colored frames cued target name type. Left: For bilinguals, English (*tree*) or Spanish (*árbol*); for monolinguals, basic-level (*tree*) or subordinate (*willow*). Right: for bilinguals, Spanish (*zapato*) or English (shoe); for monolinguals, subordinate (*sneaker*) or basic-level (*shoe*).

The number of stimuli was kept the same as that in the monolingual experiments (Experiments 3–4), which were more difficult to find. In any case, the number of stimuli used here is comparable to that used in the classical studies (ten numbers in Meuter & Allport, 1999; ten pictures in Costa & Santesteban, 2004, and Costa et al., 2006). The stimuli for the monolingual task were chosen first and replacements in the bilingual task were made to avoid cognates and low name agreement.

Design and procedure.

The pictures were presented in five blocks in a sandwich design (Figure 2) modeled after the design in Prior and Gollan (2011). At the beginning of each of the first three blocks, bilinguals read brief instructions in English about their task in the respective block, followed by practice consisting of naming all experimental pictures (once in each language in the mixed blocks). During practice, participants were informed of the intended name if they spontaneously produced a different one. The required language on each trial in the mixed blocks was indicated by the picture frame color (blue or red, counterbalanced across participants); the frames were also kept in the single-naming blocks.

In the first two pre-mix single-naming blocks (either only English or only Spanish), bilinguals named pictures on 12 practice and 36 (12 pictures × 3 repetitions) experimental trials per block. Language order was counterbalanced across blocks. A picture was never repeated on successive trials.

In the following three mixed blocks, bilinguals completed a single set of 24 practice trials followed by three sets of 48 trials of unpredictable switching between English and Spanish. Each of the three mixed blocks began with an additional dummy trial (excluded from analyses). Switch trials were half of all trials, and there were no more than four successive trials of the same type.

Language order in the two final post-mix single-naming blocks (36 trials each) was reversed from the first two single-naming blocks (A–B-mixed-B–A or B–A-mixed-A–B). The counterbalancing of language order in the pre-mix single-naming blocks (English first or Spanish first), item list order in the pre- and post-mix single-naming blocks (list order 1, 2 (pre-mix) + 3, 4 (post-mix) or list order 2, 1 (pre-mix) + 4, 3 (post-mix)), and frame color (blue or red) produced eight versions of the experiment, administered to a roughly equal number of participants (between 6 and 9). Twenty-nine bilinguals began with naming in Spanish, and 27 began with naming in English.

A trial had the following structure: A blank screen displayed for 700 ms was followed by an asterisk displayed for 700 ms, a blank screen displayed for 500 ms, and the target picture displayed for 3000 ms or until the detection of a vocal response. A trained experimenter coded different-from-intended responses and voice-key inaccuracies in real time. The DMDX software for experimental presentation (Forster & Forster, 2003) recorded participants’ vocal responses for subsequent verification.

The study procedures conformed to Federal guidelines for the protection of human subjects and were approved by the UTEP Institutional Review Board. All participants gave informed consent to participate prior to testing.

Coding and data analysis.

Analyses excluded voice key inaccuracies and experimenter errors (826 trials, or 5.1% of the data) and, subsequently, outliers – response times that were slower than 3 standard deviations above the mean or faster than 300 ms (250 trials, or 1.6% of all data). Response times on correct trials and errors (which included different words in the same language, translation equivalents, disfluencies and failures to respond; together 405 trials or 2.5% of the data) were analyzed with linear or logistic mixed effects regression modeling, respectively (LMER; Baayen, 2008; Jaeger, 2008). Error rates in all experiments are reported in Table 2, and analyses of error rates are reported in Appendix A.

Table 2.

Error rates in all experiments.

Experiment	Block type
	Single pre-mix	Mixed Stay	Mixing cost	Switch	Switching cost	Single post-mix	Post-mixing cost
Experiment 1
English	2.04	2.72	.69	3.49	.77	.65	−1.39
Spanish	3.13	2.71	−.41	4.94	2.22	.70	−2.43
Difference	1.09	−.01		1.45		.05
Experiment 2
English (dom.)	2.50	2.94	.44	3.60	.66	1.17	−1.33
Spanish (nondom.)	8.29	4.82	−3.47	6.29	1.47	2.98	−5.31
Difference	5.80	1.88		2.69		1.81
Experiment 3
Basic-level (dom.)	2.49	3.35	.86	4.29	.94	1.74	−.75
subordinate (nondom.)	1.85	2.05	.20	4.05	1.99	.75	−1.10
Difference	−.64	−1.29		−.24		−.98
Experiment 4
Basic-level (dom.)	1.45	2.93	1.49	3.27	.34	1.85	.41
subordinate (nondom.)	3.30	2.00	−1.30	3.51	1.52	1.45	−1.85
Difference	1.85	−.94		.24		−.41

Open in a new tab

Switching costs were assessed with a model comparing stay and switch trials in mixed blocks. The fixed predictors in this model were Switching trial type (stay trials, coded as −0.5, switch trials, coded as 0.5; all other trials were coded as NA), Language (English, coded as −0.5, Spanish, coded as 0.5) and their interaction. Mixing costs were assessed with a model comparing pre-mix single-naming trials and stay trials in mixed blocks. The fixed predictors in this model were Mixing trial type (pre-mix trials, coded as −0.5, stay trials, coded as 0.5; all other trials were coded as NA), Language (coded as above) and their interaction. Post-mixing costs were assessed with a model comparing the pre-mix and post-mix single-naming blocks. The fixed predictors in this model were Single naming block type (pre-mix block, coded as −0.5, post-mix block, coded as 0.5; the mixed block was coded as NA), Language (coded as above) and their interaction. Repetition benefits for each picture name in each language were assessed in a model with the fixed predictors Block type (single, coded as 0.5, and mixed, coded as −0.5), Language (coded as above) and Repetition (a continuous mean-centered predictor). Language dominance effects were assessed in all models with the Language predictor. Additional supplementary analyses, where pertinent, are described in the Results section.

All models were implemented with the maximal random-effects structure justified by the design (Barr, Levy, Scheepers, & Tily, 2013) using the lmer or glmer functions in the lmerTest package (version 3.1–0, lme4 version 1.1–21) in R (version 3.6.1), using the Satterthwaite method to approximate denominator degrees of freedom for models with a continuous dependent variable. If the full random-effects model did not converge, the model was simplified by first removing random-effects correlations, and then removing step-wise the random effects accounting for least variance, with the restriction that random slopes were removed before random intercepts. The bobyqa optimizer was used to facilitate convergence.

Trial-level data for all experiments are publicly available at https://osf.io/67yad/.

Results

The results of all statistical models are reported in Table 3. By-participant mean naming latencies are plotted in Figure 3 (per block type and language) and Figure 4 (per repetition in initial and final blocked naming).

Table 3.

LMER analyses of the data of Experiment 1

Model	Predictors	Estimate	SE	t or z	p
Switching costs
	Switching trial type (stay vs. switch)	46.00	8.40	5.48	< .001
	Language	16.45	16.70	.99	.34
	Switching trial type × Language	−15.34	10.04	−1.53	.13
Mixing costs
	Mixing trial type (pre-mix vs. mixed-stay)	198.10	14.57	13.59	< .001
	Language	33.70	17.72	1.90	.07
	Mixing trial type × Language	−18.14	18.37	−.98	.33
Post-mixing costs
	Single naming block type (pre-mix vs. post-mix)	66.94	12.56	5.33	< .001
	Language	29.06	15.78	1.84	.08
	Single naming block type × Language	−27.82	14.95	−1.86	.07
Repetitioin benefits
	Block type (single vs. mixed)	−206.84	13.20	−15.67	< .001
	Language	21.30	15.18	1.40	.17
	Repetition (single: 1–3; mixed: 1–6)	−14.85	2.72	−5.46	< .001
	Block type × Language	2.90	10.41	.28	.78
	Block type × Repetition	−15.91	5.11	−3.11	.009
	Language × Repetition	−5.43	3.48	−1.56	.14
	Block type × Language × Repetition	−6.36	7.20	−.88	.39

Open in a new tab

Note: Dark grey shading indicates significant effects, light grey shading indicates marginal effects.

Figure 3. — By-participant mean naming latencies for each block type and language for balanced bilinguals in Experiment 1. Error bars represent 95% confidence intervals.

Figure 4. — By-participant mean naming latencies for each repetition in initial and final blocked naming for balanced bilinguals in Experiment 1. Error bars represent 95% confidence intervals.

Switching costs.

There were significant switching costs: Response times were slower on switch than on stay trials in mixed blocks (Switching trial type was a significant predictor). In other words, balanced bilinguals named pictures more slowly when the immediately preceding picture was in a different language than when the immediately preceding picture was in the same language. As predicted, these costs did not differ across balanced bilinguals’ languages (the interaction between Switching trial type and Language was not significant).

Mixing costs.

The pattern for mixing costs was the same as for switching costs: There were significant mixing costs such that response times were slower on stay trials in mixed blocks than in pre-mix blocks (Mixing trial type was a significant predictor). In other words, balanced bilinguals named pictures in the same language as immediately preceding pictures more slowly within a block with frequent switches than within an initial block with no switches. As predicted, the mixing costs were equivalent for balanced bilinguals’ two languages, i.e., symmetrical (the interaction between Mixing trial type and Language was not a significant predictor).

Post-mixing costs.

Blocked naming after trial-level language switching also incurred a cost: Naming latencies in the post-mix single-naming blocks were slower than those in the pre-mix single-naming blocks (Single naming block type was a significant predictor). There was a trend for larger post-mixing costs for English than for Spanish (the interaction between Single naming block type and Language was marginally significant).

Repetition benefits.

As expected, naming latencies got faster with repetition (Repetition was a significant predictor). This improvement was larger in single-naming blocks than in mixed blocks (the interaction between Block type and repetition was significant). To shed light on this interaction, we ran additional models on the separate data of the pre-mix, mixed and post-mix blocks with Language, Repetition and their interaction as fixed predictors. These analyses showed significant repetition benefits across all blocks (pre-mix block [Repetition predictor: Estimate = −20.90, SE = 4.80, t = −4.36, p < .001]; mixed block [Repetition predictor: Estimate = −6.99, SE = 2.34, t = −2.99, p = .007]; post-mix block [Repetition predictor: Estimate = −25.03, SE = 8.63, t = −2.90, p = .01]). Further, the repetition benefits were larger in Spanish than in English in the pre-mix block [Repetition × Language interaction: Estimate = −12.06, SE = 5.65, t = −2.13, p = .04], but not in the post-mix or mixed blocks [for the Repetition × Language interaction, both ps > .6].

Language dominance effects.

Balanced bilinguals named pictures with a roughly similar speed in English and Spanish: The Language predictor was marginal in the mixing and post-mixing costs analyses (a trend for faster naming in English), and non-significant in the switching costs and repetition benefits analyses. Separate models on the data of the pre-mix and post-mix blocks aimed to see how the trend for faster naming in English developed in the course of the experiment. These models (with Language as the only fixed predictor) showed that the language difference was significant in the pre-mix blocks [Estimate = 43.22, SE = 19.50, t = 2.22, p = .04] but was not significant in the post-mix blocks [Estimate = 15.12, SE = 15.15, t = 1.00, p = .33]. In other words, there was an advantage of English over Spanish in the beginning of the experiment, but it disappeared by the end of the experiment. The marginally larger post-mixing costs for English (72 ms.) than for Spanish (49 ms.) indicated that this was at the expense of English (see Figure 12).

Figure 12. — Magnitude of the advantage gained by nondominant names over dominant names from initial to final blocked naming in all experiments. It was calculated as (dominant minus nondominant naming latencies in the pre-mix blocks) minus (dominant minus nondominant naming latencies in the post-mix blocks).

Discussion

Balanced bilinguals in Experiment 1 showed significant switching and mixing costs, which, as predicted, did not differ across their two languages. Assuming that language control mechanisms act to allow less effortful production in the nondominant language, it is expected that similar levels of control would be applied to two languages of similar dominance because neither of their languages is much more likely to interfere than the other.

Some features of the data, however, indicated that the two languages of these bilinguals were not represented completely equivalently. There were larger repetition benefits for Spanish at the beginning of the experiment as well as a reduced difference between the languages at the end of the experiment relative to the beginning, at the expense of English. In view of the accepted interpretations of such effects in prior literature, both of these results suggest a slight dominance of English. This is in accordance with these bilinguals’ language history reports, in which English (with an average self-rated proficiency of 9) was rated as slightly more dominant than Spanish (with an average self-rated proficiency of 8.5 [t(55) = 2.34, p = .02]). However, dominant deprioritization effects did not extend to a reversed dominance effect (unlike in Costa et al., 2006; Declerck et al., 2020b; Kleinman & Gollan, 2016; Li & Gollan, 2018): Spanish was produced more slowly in initial blocked naming and with a similar speed to English in the subsequent blocks. This might be because our Spanish words were longer and of lower frequency than the English words.

Repetition benefits were asymmetrical at the beginning of the experiment but equivalent thereafter. Our interpretation of this finding is that the baseline accessibility of the names in the two languages was slightly different but it became more similar during their repeated retrieval in a mixed context. This pattern is consistent with the names’ selection thresholds being adjusted in a mixed context (through inhibiting the more dominant one more strongly or through adjusting one or both thresholds directly).