Speech production is inherently guided by top-down control processes—the speech act is initiated by the speaker’s intention to communicate, and speakers monitor their output to ensure alignment with their interlocutors (Levelt 1989
However, during speech, numerous possibilities arise for conflict between competing representations, both within a language (Andrews 1997
) and, in the case of multilinguals, between languages (Costa et al. 1999
; Hermans 2000
; Hermans et al. 1998
; Kroll et al. 2006
; Schwieter and Sunderman 2009
). Thus, at multiple levels within the language system, there is a need for cognitive control mechanisms to support multilingual speech production. The extent to which language control for bilinguals overlaps with other non-linguistic processing has been debated for years (e.g., Branzi et al. 2016
; Calabria et al. 2019
; Declerck et al. 2017
; Segal et al. 2019
). Although behavioral studies of multilingual language control have separately examined the importance of different executive functions (EFs), including inhibitory control (IC; Linck et al. 2012
; Koch et al. 2010
; see also Antoniou 2019
; Bialystok 2017
), working memory (WM; e.g., Christoffels et al. 2003
; see also a meta-analysis by Grundy and Timmer 2017
), and task-switching (Prior and Gollan 2011
; Timmer et al. 2019
; Wiseheart et al. 2016
), little is known about the joint contributions of the collection of EFs. In this study, these three EFs were examined simultaneously to estimate their roles during a language-switching task that placed specific demands on language control functions.
Cognitive control has been argued to operate by various mechanisms. On the one hand, global or sustained control guides behavior by keeping active the current goal of the system (Braver et al. 2003
). Such global control is necessary to ensure the goal-appropriate response is selected, particularly when multiple task-relevant responses are available for response selection. However, even with a clear goal in place, competition occurs in many forms, such as between distractor and target representations, between competing responses, or between different dimensions of the target representation (see the Dimensional Overlap Model, Kornblum et al. 1990
). With language tasks, even more opportunities for conflict appear throughout the system at the phonological, orthographic, and morphosyntactic levels (Kroll et al. 2006
). Some additional local control process(es) must be engaged to facilitate the selection and execution of the correct task-relevant response.
This distinction between global and local control processes has been discussed in the literature on bilingual language control (e.g., De Groot and Christoffels 2006
) and fits well with models of bilingual language processing. In Green
) IC Model, language task schemas provide sustained control by orienting the system towards performing the goal-relevant task, and potential responses that conflict with the current goals of the system are inhibited to prevent errors. The top-down control of the language task schemas and the reactive inhibitory mechanism work together to resolve cross-language conflict between representations in the two languages. The revised Bilingual Interactive Activation (BIA+) model of word recognition (Dijkstra and Van Heuven 2002
) also includes task schema and inhibitory mechanisms. Neurophysiologically motivated models have postulated the importance of the frontal regions—including the anterior cingulate cortex and dorsolateral prefrontal cortex—to language control (Abutalebi and Green 2008
), implicating domain-general EFs that support cognition more broadly (see also Levy and Anderson 2002
Individual differences in global/sustained and local control likely contribute to multilingual speech production in different ways. In Green
) IC Model, sustained control may be supported by the maintenance of goal representations in WM—a critical component to the control of a range of goal-directed behaviors (e.g., Engle 2002
). Thus, individuals with larger WM capacity may better engage global control. In contrast, local control may be enacted by engaging inhibitory mechanisms, such as the reactive inhibition of representations that are competing with the target representation, as has also been posited by models of memory retrieval (e.g., Levy and Anderson 2002
). Green’s IC Model focused on reactive inhibition that is triggered by the activation of non-target representations. Colzato et al.
) identified a two-component model of cognitive control that included both a global control mechanism and a separate IC component.2
We build upon these theoretical frameworks to examine the contributions of different EFs to language control during a trilingual speech production task.
Recent studies have also shown a link between bilingual language control and domain-general executive control by assessing the effect of language-switching training (Liu et al. 2019
; Prior and Gollan 2013
; Timmer et al. 2019
). Although Timmer et al. found that training in language switching transfers to the non-linguistic domain for certain sub-mechanisms (i.e., switch cost) but not for others (mixing cost), Prior and Gollan reported no transfer effects, neither for switch cost nor mixing cost. Liu et al.’s study found that training in language switching reduced mixing costs and the anti-saccade effect among bilinguals. They argue that extensive training in monitoring and inhibitory control enhances the corresponding components of cognitive control.
IC and WM seem particularly relevant to language control during language switching. A recent study (Kaushanskaya and Crespo 2019
) found that the effects of exposure to code-switching input were modulated by WM. For children with high WM, being exposed to code-switching input did not negatively affect their language skills. However, in both receptive and expressive language skills, children with low WM were negatively affected if they were exposed to code-switching input. Thus, WM can and may have a modulating effect on IC. Linck et al.
) found that better inhibitors showed smaller switch costs in a trilingual language-switching task. In other words, more efficient IC skills resulted in smaller switch costs into a non-target/irrelevant language. Moreover, previous work has shown that language switching led to increased activation of the DLPFC—specifically in areas previously linked to IC mechanisms responsible for resolving conflict (Hernandez et al. 2000
; for a review, see Abutalebi and Green 2008
). Additional work has demonstrated enhanced general executive functioning among bilinguals exhibiting better language control on a range of cognitive control tasks requiring attentional control (Festman et al. 2010
) and conflict resolution (Festman and Münte 2012
). Other works (e.g., Abutalebi et al. 2001
; Hernandez 2009
; Price et al. 1999
) have used neuroimaging studies to examine the broader basis of language switching. What is still unclear, however, is how WM can interact with EFs and IC.
Asymmetries in language dominance will also constrain the effect of IC on switch cost. We should see the largest effects when switching from the least dominant language (L3) to the most dominant language (L1) (Green 1998
). This would be consistent with Meuter and Allport
) as well as Linck et al.
) findings that there is a cost to re-engage the previously irrelevant L1. Differences in dominance also have neural consequences. Recent work (Garbin et al. 2011
) found that language switching is constrained with proficiency; early and high-proficient bilinguals implement different brain networks than low proficient bilinguals. For low-proficient or unbalanced multilinguals, naming in the least dominant language is expected to place the largest demands on working memory. In order to produce in the weak L3, or even the weak second language (L2) in the case of this study, the individuals must suppress the dominant L1 and focus their attention on the weaker languages.
Our research question in the current study is: Is there a relationship between certain EFs—specifically working memory updating, inhibitory control, and task-set switching—and language control? Given the findings above, one might expect the abbreviated answer to this question to be “yes”. If IC is engaged to reduce cross-language representational conflict during speech production in a mixed language context, then IC should be most relevant to performance in conditions where the greatest amount of cross-language interference is expected, namely when switching into or out of L1. Specifically, we predicted that better IC abilities should allow more efficient deployment of inhibition in the face of conflict that arises when switching into a previously irrelevant language, and thus should be related to smaller switch costs, due to faster latencies in switch trials (Linck et al. 2012
). In contrast, because L1 is the dominant response language, there should be greater demands on WM resources to activate and maintain the task schema for L2 or L3 naming relative to L1 naming, and thus WM should be most relevant when naming in the less dominant languages. Specifically, more efficient activation and sustainment of the L2 naming and L3 naming task schemas should facilitate switching into L2 and L3, leading to reduced switch costs. Furthermore, following Prior and Gollan
), we expect that better task switchers (i.e., those exhibiting smaller switch costs on a monolingual task-switching task) will be better language switchers in mixed language contexts due to the presumed shared cognitive processes.
This study was designed to examine the relationship between different EFs and language control during trilingual speech production. We measured three EFs—inhibitory control, working memory updating, and task-set switching—that are related but separable subcomponents of the executive control system (Friedman and Miyake 2004
) and have been linked to bilingual language processing and implicated in studies on the cognitive benefits of bilingualism. Although previous research on multilingual speech production has studied the individual role of these EFs in isolation, this is the first study to simultaneously assess their relative contributions to language control.
The effect of IC on switch costs was constrained to precisely the condition where the asymmetry in language dominance should induce the largest effects of lingering inhibition—when switching from the least dominant L3 to the most dominant L1 (Green 1998
). These results are consistent with recent claims that better inhibitors can more efficiently re-engage the previously irrelevant L1 (Linck et al. 2012
), although the within-condition slopes suggest a further specification of this account. The switch cost-IC effect appears to have been driven by overall slower access to L1, combined with an increase in control on switches. Slowed L1 lexical access following naming in a less dominant language has been found even after a delay of 10–15 min (see Levy et al. 2007
). In their study, Levy et al. similarly asked bilingual learners to switch between languages when naming pictures, but later cued the retrieval of the L1 names of those pictures with a novel phonological rhyme cue (e.g., “break—s___” to cue the retrieval of snake). For pictures that had repeatedly been named in L2, participants were less likely to successfully retrieve the L1 name in response to this rhyme cue. Based on the extensive literature on retrieval-induced forgetting that has implicated inhibitory mechanisms in the control of memory retrieval (see Anderson 2003
, for a review), Levy et al. concluded that the forgetting effects indicated that L1 lexical access had been impaired by the lingering inhibition of the L1 picture names.6
In the present study, perhaps better inhibitors were able to more efficiently engage IC mechanisms to support retrieval in the weaker L3, manifesting as slowed re-engagement of L1 in switch trials relative to non-switch trials (in which the activation of L1 has rebounded).
For unbalanced multilinguals, naming in the least dominant language is expected to place the largest demands on working memory, since the speaker must focus attention on the weakest language in the face of distracting interference from the dominant L1 (and in the case of the present study, another weak language). One useful strategy could be to always prepare not to use L1. This strategy would clearly benefit L3 naming, where L1 is most likely to interfere. Yet this strategy would potentially impair L1 naming trials. This may have the unexpected consequence of increasing demands on working memory to engage the L1 naming schema, despite the relative dominance of L1.
Indeed, this is where WM updating effects were found: better WM updating was related to larger switch costs when switching into or out of the dominant L1. This is a somewhat counterintuitive finding—better WM is typically related to better, not worse, performance. However, the WM effects were driven by a relationship between better WM and faster latencies in non-switch trials (when the same task set must be maintained across trials) but no differences appeared in switch trials. Perhaps in this mixed language naming task, the better updaters were able to engage a top-down strategy of globally biasing against responding in L1 during L3 naming, and this global bias spilled over momentarily upon switching into L1. However, then, the better updaters were able to quickly re-engage L1 and disengage the “inhibit L1” task schema, leading to faster responses on L1 non-switches. In the same vein, Kroll et al.
) found a counterintuitive WM result—individuals with better WM resources showed smaller cognate facilitation effects. This finding suggests that the participants were able to focus their attention on L2 to avoid any potential interference from L1, which, for cognates, prevented them from benefiting from the L1 overlap (in orthographic and phonological form). In a parallel manner, the present data suggest that better WM resources may have allowed participants to focus attention away from L1 when using L3, and the effects of this strategy of pushing L1 out of the focus of attention lingered when switching back into L1.
Taken together, the WM and IC results of the current study suggest that language switching may reflect at least two components: a repetition benefit (modulated by available attentional resources), and a switch cost (modulated by IC abilities). Our WM results suggest that individuals with greater WM resources benefit more from task repetitions, since better working predicted faster latencies in non-switch trials. This is consistent with recent claims that switching effects in (monolingual) task-switching reflect a repetition-induced benefit to performance, rather than a switch-induced cost to performance (De Baene et al. 2012
). However, the IC results seem to fit with an inhibition-based account of task-switching (Mayr and Keele 2000
). In our study, better inhibitors benefited less from task repetitions and, critically, showed reduced costs due to switching. That is, across participants, there was a cost to performance in switch trials, but more efficient inhibitors suffered less of a cost. Moreover, this IC effect was independent of any WM effects, suggesting that inhibitory mechanisms also support language switching. To our knowledge, this study is the first to simultaneously model and examine the effects of multiple EFs on language and task-switching performance.
These accounts are both congruent with existing claims that the relative dominance of the two languages determines the extent to which cognitive control is required for successful language selection (Green 1998
; Meuter and Allport 1999
). Indeed, when these participants switched between less asymmetric language pairs (e.g., between L2 and L3), no reliable effects of the EFs were found, replicating our previous pattern of results (Linck et al. 2012
). That is, cognitive control effects are constrained to conditions in which the greatest amount of competition or response conflict is expected, much as the cognitive benefits of bilingualism most reliably emerge in conditions high in conflict and/or cognitive load (Bialystok and Craik 2010
; Hilchey and Klein 2011
Finally, in terms of task-switching, our results indicated that this was not related to switch costs in any language, but we did find that better task switchers were faster to name pictures in L2. Given the mixed language context of the task and the demands to shift quickly between languages, we find this result to be in line with Festman and Münte
) who found that the individuals they had classified as ‘non-switchers’ (individuals who are able to avoid unintentional language switches) outperformed ‘switchers’ (those individuals who are less able to avoid unintentional language switches) in terms of speed and accuracy on a series of EF tasks. Festman and Münte cite this as evidence for a relationship between bilingual switching behavior and general cognitive control, a claim our data support as well.
Integrating Models of General Cognitive Control and Multilingual Language Control
Research on the cognitive control of attention and memory systems has generated a theoretical framework that implicates both global/sustained control mechanisms and local control mechanisms (Braver et al. 2003
). This framework has broad application to a range of cognitive tasks involving distractor or response conflict (e.g., Flanker task, Stroop task) and ambiguity resolution (e.g., during sentence comprehension), and also can be easily incorporated into existing models of bilingual language processing. Green
) IC Model posits global control via task schemas, which are activated based on the current goals of the speaker. This global control mechanism provides top-down ‘supervisory’ control over the system through the task schemas. When non-target language representations that conflict with the active task schema are activated, their production is prevented by localized inhibition that is applied to the specific representations based on their degree of competition with the target representation and/or task schema. Inhibition might also be applied globally to an entire language, and indeed evidence of both global and local inhibition effects has been reported (Guo et al. 2011
). In the current study, WM effects on switch costs were constrained to non-switch trials, where successful performance depends on maintaining active the current response set/task schema, whereas IC effects on switch costs were found to impact both non-switch and switch trials, with good inhibitors being faster in general, but even more so in switch trials. This suggests that IC supported performance across the board, and especially in switch trials. WM was also important, particularly in non-switch trials when needing to maintain the current response set.
The finding that WM and IC independently accounted for variability in switching performance leads us to speculate that the IC Model could be further specified to incorporate these two EFs at different levels. WM updating effects might operate at the level of the language task schemas to provide top-down guidance of the selection process. In contrast, IC effects might be localized to the level of individual lexicosemantic representations, as currently suggested by the IC Model (see also Levy et al. 2007
, for similar claims motivated by the retrieval-induced forgetting literature). These two levels are necessarily interrelated: the inhibition of specific representations is motivated/cued, at least in part, by the top-down guidance from the language task schemas, whose job it is to bias production towards the target language (Green 1998
These modifications to the IC Model would incorporate theoretical claims from the literature on cognitive control, while also extending the IC Model to make predictions about the impact of individual differences in EFs. This approach provides a unified account of results from this study and other recent examinations of individual differences in language control (Festman and Münte 2012
; Festman et al. 2010
; Prior and Gollan 2011
). Our results build on these findings to provide more specified links between executive functioning and online language processing.