Trajectories to third-language proficiency

Introduction

Conversation is a primary site for language use and the ability to converse with others is a marker of communicative proficiency in the language of use. Pragmatic skills are also essential to everyday conversation. In the case of adult, proficient bilingual speakers we can assume that these operate satisfactorily and so consider other aspects of the trajectories to communicative proficiency in a third language (L3). In this theoretical paper, we focus on the processes of language control involved in representing and using grammatical forms. Individual differences in these processes we suggest give rise to different trajectories to L3 proficiency and such differences operate even in the initial stages of acquisition.

Conversation is a form of communicative action and, as in other forms of action, there are cycles of action and perception that involve multiple levels of control (see Figure 1). For the present purposes we distinguish two levels of control that work together to coordinate conversational exchanges. To achieve communicative goals, utterances are guided by top-down processes of control that activate relevant lexical concepts and grammatical forms, and inhibit competing ones. Bottom-up processes of control elicited by utterances of a conversational partner, activate a network of existing lexical concepts and grammatical forms or trigger inductive processes to account for their content. In learning the grammatical forms of a L3, different types of input (formal instruction, heard speech, transfer) load differently on these processes. So, for example, formal instruction in the syntactic forms in a L3 loads on processes of top-down control to alter representations. By contrast, participating in a conversation with a L3 speaker can trigger bottom-up control processes that seek to account for the input in terms of known structures, or, where these are insufficient, lead to the induction of novel syntactic forms. Transfer of grammatical forms from one or other of a bilingual speaker’s prior languages is likely to play a prominent role in learning a L3, at least in the initial stages. In addition, transfer also requires processes of language control. Current theoretical models of transfer make strong claims about the nature of initial transfer but leave to one side detailed discussion of language control demands that transfer requires. We consider three models of transfer and explore their control demands in the section entitled “Three models of transfer and the demands on language control”. We argue that the syntactic landscape during the initial stages of L3 learning is likely to be more varied than allowed for by current models of transfer. The landscape, we suggest, will reflect the outcome of both top-down and bottom-up control processes. Individual differences in implementing these control processes will yield different starting points for trajectories to L3 proficiency.

OPEN IN VIEWER

Figure 1.

Sketch of the language control schemas and the language network.

In order to use syntactic forms (whether transferred or induced) speakers need to represent them in the context of their prior mental representations of the syntactic forms of their first language (L1) and second language (L2). We consider what is involved in reshaping the landscape of syntactic forms in the section on “Representational change in L3”. We identify the changes required (both initially and with increased proficiency) by extending the bilingual network model proposed by Hartsuiker, Pickering, and Veltkamp (2004). Change in the language network is a necessary but not a sufficient condition to allow speakers to select the relevant syntactic form for L3 production. Experimental data indicate that language control mechanisms external to the language network are required. Such mechanisms, organised to control two languages, need to adapt to meet the demands of controlling three languages. We consider the changes that are required to meet these novel demands in the section “Adapting the control structure for language selection in the L3”. Adapting effectively is likely to be constrained by cognitive differences amongst speakers and also by how they actually used their two prior languages. Speakers may differ, for example, in whether they used their two prior languages in distinct conversational contexts or whether they code-switched between their two languages within a conversational context. Any adaptive change to meet novel control demands will also have consequences for how speakers control their prior languages. It follows that developing a fuller understanding of the trajectories to L3 proficiency also involves understanding correlating effects on the production of the L1 and L2. The “Review and future directions” section summarises the main arguments and makes further suggestions for exploring individual differences in the trajectories to L3 proficiency and precedes a concluding statement.

Three models of transfer and the demands on language control

Three prominent theoretical models propose different sources of language transfer in the initial stages of L3 language learning but in one way they are alike: they share the presumption that there is a single trajectory to the L3 in the initial stages: it is the trajectory proposed by the model in question. In the L2 status factor model (e.g. Bardel & Falk, 2007, 2012) transfer depends on the nature of the mediating memory representations. L3 learning is held to use a declarative representation of grammar and such a representation is more accessible for the speaker’s L2 than for their L1 (cf. Paradis, 2009; Ullman, 2001). Transfer in this account involves structures from the source language (L2) and none from the non-source language (L1). ¹ Transfer is also held to be all-or-none and so transfer from the L2 can be facilitative with respect to the L3 or it can be non-facilitative (i.e. involve a structure that is in fact not part of the L3 grammar). In the cumulative enhancement model (Flynn, Foley, & Vinnitskaya, 2004), by contrast, transfer arises from either of a speaker’s languages regardless of their order of acquisition. Also, in contrast to the L2 status factor model, transfer can only be facilitative. In the typological primacy model, TPM for short (e.g. Rothman, 2011, 2015; see also Rothman, Alemán Bañón, & González Alonso, 2015), as in the cumulative enhancement model, grammatical representations of both languages are possible sources of transfer independent of their order of acquisition but in contrast to that model actual transfer is determined by which language is considered to be most structurally similar to the L3 based on limited L3 input. Also in contrast to the cumulative enhancement model, but in common with the L2 status factor model, transfer from the source language can be non-facilitative. ² In other words, for both the TPM and the L2 status factor models transfer is all-or-none.

These alternative models articulate three distinct learning trajectories at the initial stage of L3 learning. Experimentally this is excellent because it permits decisive tests: if there is a single trajectory then one model can be correct and the others false. However, suppose the presumption of a single trajectory during initial L3 learning is false. This means that variability in the trajectory to L3 proficiency is present from the start and all these models would need to be revised. In some circumstances, as we will see, the precise nature of the variability undermines the predictions of the models.

Are there any grounds for supposing variability in the initial trajectories? We consider three. The first derives from a logical necessity and intimates that variability is inherent. The argument runs as follows: even if an important goal is to understand the initial stages of L3 acquisition, ultimately we need to understand how individuals become more or less proficient speakers of the L3. All-or-none transfer in the initial stages implies that syntactic forms from the source language populate the L3 landscape in their entirety. Yet if negative transfer (i.e. the transfer of non-facilitatory syntactic forms) is to be overcome, these forms must be potentially dissociable. If this were not the case, rejection of one structure would involve rejection of all and render transfer superfluous. The simplest solution to overcoming negative transfer is to suppose that syntactic forms are dissociable from the start: in other words, transfer is the transfer of dissociable syntactic forms. The cumulative enhancement model necessarily envisages this type of transfer because it distinguishes transfer appropriate and transfer inappropriate forms.

Granting dissociability allows a probabilistic element into initial transfer and is the basis of a second ground for supposing that variability is inherent to initial transfer. Dissociability does not mean that all-or-none transfer is a theoretical impossibility but it does mean that we need to distinguish between the factors that shape the transfer landscape as a whole and the particular structures that populate it. We can ask which factors influence the likelihood of transfer of a given structure. A speaker’s proficiency, experience and use of their two prior languages will all be relevant. Evidently low proficiency may mean that a specific grammatical form is not part of the speaker’s knowledge representation for that language and so that form is unavailable for transfer. Tests of models of initial transfer in the L3 quite rightly ensure their bilingual participants are highly proficient speakers of both their languages. However, speakers will use a given form with varying frequency that may affect its availability for transfer. For proficient L2 speakers, extensive use of two languages may affect properties of the representations that are open for transfer because the relationship between the L1 and L2 is bidirectional: language systems are permeable (Kroll, Dussias, Bice, & Perrotti, 2015). For example, processing preferences in the L2 are known to alter the way speakers resolve syntactically ambiguous relative clauses in their L1 (Dussias & Sagarra, 2007). Native Spanish speakers reading the Spanish translation of “Someone shot the son of the actress who was on the balcony” typically name “the son” in response to the question “who was on the balcony?” and so reveal a high attachment preference. Native English speakers by contrast, typically respond “the actress” and so reveal a low attachment preference. The critical result for the present purposes is that Spanish-English bilinguals living in an English-speaking environment preferred the low attachment strategy when reading in their first, and native, language Spanish. The extent of such permeability is unknown but current research strongly suggests dependence on the patterns of language use (Valdés Kroff, 2012).

A third ground directly concerns the process of language control. The fundamental claim of the TPM and L2 status factor model is that only one source language shapes and populates the initial landscape of transfer even if the syntactic forms in that source language have been influenced by processing of structures in the non-source language. This claim implies a language control constraint: namely, that the non-source language plays no role in shaping or forming the initial landscape of syntactic forms in the L3. Now such a constraint operates in a specific mental environment and implies a top-down process of language control. A prerequisite for transfer is that the person’s other languages have to be accessible. We know from a range of experimental studies that in the normal course of events both languages are active in bilingual speakers even when they intend to use just one of them (Kroll, Bobb, & Wodniecka, 2006). Transfer then for the TPM and L2 status factor model must involve a process of top-down control in which one language is selected for transfer (on grounds specific to the model) and the other language is completely inhibited from playing any role. What type of language control process is consistent with the proposal of the cumulative enhancement model in which transfer is only facilitative? Suppose a top-down control process. Such a process could work if the mental system were to have full knowledge of the syntactic forms required in a L3. In this case, any appropriate pre-existing identical forms could be recruited via a top-down control process. However, this proposal merely raises a further question: what is the source of such knowledge? A better alternative, and one in line with the proposal of the cumulative enhancement model that structures are “searched for” (Flynn et al., 2004; see Bardel & Falk, 2007), is that transfer under this model is the outcome of bottom-up control processes. Such processes seek to account for activation triggered by speech input. Pre-existing structures in either language may do so and such structures are necessarily facilitative. Because of the intimate connection between the processes of speech comprehension and production (Pickering & Garrod, 2013; Silbert, Honey, Simony, Poeppel, & Hasson, 2014), activation of structures by speech input can make these structures available for speech output. Figure 1 illustrates the relationship between the language control processes and the language network that undergoes change.

To summarise so far: in a two-dimensional control space, where one axis refers to top-down control and the other to bottom-up control, the TPM and the L2 status factor model occupy one quadrant of the control space: high-top-down control, low bottom-up control. By contrast, the cumulative enhancement model occupies a diagonally opposite position: high bottom-up control, low top-down control. If both classes of processes are at work then the initial landscape of transfer will be much more dynamic than that predicted by any of the transfer models. For example, the overall landscape may be shaped top-down by one source language but facilitative structures from the non-source language may be represented as an outcome of bottom-up processes triggered by speech input in the L3. Individuals may also differ in their initial positions within this space. For example, a speaker’s skills in top-down control will affect their ability to restrict the transfer of structures from a single source language. The weighting of these two classes of process in shaping and populating the L3 landscape is likely to change over the course of L3 learning. Immersion in a L3 environment, for instance, will increase the relevance of results from bottom-up control processes. Here too, individual differences will continue to play a role as speakers will differ in the facility with which they infer novel structures rather than use contextual/pragmatic information to infer intended meaning.

Representational change in the third language

Granted the parallel activation of languages during speech production, speech production in the L3 is a competitive process in which lexical items tied to the source language must be blocked from production. It follows that successful transfer requires a context-sensitive resolution of competition. For this to take place syntactic forms for L3 use must be identified along with L3 lexical items. Such identification involves a representational change. In the next section we consider the nature of this change both for initial transfer and for overcoming negative transfer. In the following section, we briefly consider the sources of information that may lead speakers to overcome negative transfer.

Transfer and the language network

To visualise the representational change required, consider a language network in which lexical concepts map onto lemmas and word forms (Levelt, Roelofs, & Meyer, 1999). Within a language network for bilingual speakers, items and constructions for each language can be distinguished through connections to distinct language nodes, one for the L1 and one for the L2 (see, for example, Hartsuiker et al., 2004). These connections effectively tag lexical items and constructions according to language (see De Bot & Jaensch, 2015, for an alternative view). More abstract “combinatorial nodes” capture common syntactic patterns or frames such as active and passive constructions. Where two languages share a structure the relevant combinatorial node is linked to two language nodes. For example, in the case of Spanish and English (examples from Hartsuiker et al., 2004), the combinatorial node for the active construction (as in the Spanish: “El taxi persigue el camion” and the English: “The taxi chases the truck”) would be linked to a Spanish language node and an English language node.

We can extend this network view to cover initial transfer for the L3 (see Figure 2). In network terms, transfer from a single source language involves creating a language node for the L3 (in addition to that for the source language) and projecting connections from the combinatorial nodes of the source language (the syntactic frames) to the L3 language node (or vice versa). As proposed above, empirical factors (e.g. frequency of use of a syntactic structure) may constrain precisely which structures populate that landscape. In other words, we can say that initial transfer warps or shapes the L3 landscape as a whole and allows a probabilistic representation of particular structures or syntactic frames. Actual speech necessarily involves other component processes, including ones in which the slots in any transferred syntactic frame are filled with lexical items (lemmas) that are also tagged as L3.

OPEN IN VIEWER

Figure 2.

A fragment of the language network.

If, as the TPM and L2 status factor model predict, initial transfer involves the transfer of non-facilitative forms (i.e. involves negative transfer) an increase in L3 proficiency requires a further process of representational change in which these structures are no longer treated as structures suited for use in the L3. How might the network change? One possibility is that the connection to the L3 language node from the combinatorial node for that structure is removed, leaving only the connection to the language node for the source language. Empirically, it would be as if the transfer of this structure had never happened. An alternative possibility is that the connection between the combinatorial node and the L3 language node remains but any activation of that combinatorial node is inhibited in a L3 context, preventing use of that structure in overt speech (see Figure 2). Empirically, this alternative means that the network preserves the history of changes to it. If this is the case, the preceding state of the network, in which an incorrect form is tagged for L3 use, might be unmasked under certain conditions. For example, under cognitive stress, its activation might be detected in the speech of proficient L3 speakers.

Given the permeability of language representations, understanding changes in L3 proficiency associated with the elimination of negative transfer also requires us to explore changes in the production of the L1 and L2. Cabrelli-Amaro, Amaro, and Rothman (2015) report an intriguing cross-sectional study that attests to such changes with respect to a particular type of transfer error. We reference this study to illustrate two points: a theoretical point about variety in the initial transfer landscape and a methodological point about the need for converging methods.

Cabrelli-Amaro et al. (2015) examined Subject to Subject (S-to-S) raising over a dative experiencer in Brazilian Portuguese (BP) as a L3. English allows such a construction (e.g. “John seems to me to be a nice person”) and so does BP (“O Joao me parece ser uma pessoa excelente”). Spanish, however, does not (* “Juan me parece ser una persona excelente”). On the basis of overall structural similarity Spanish is closer than English to BP. Based on a scalar judgement task, and in line with the TPM, both groups rejected the acceptability of BP sentences with S-to-S raising over a dative experiencer consistent with the idea that this structure was absent from the landscape of transfer – a type of negative transfer. By contrast, an advanced and immersed group of L3 learners (native English speakers highly proficient in Spanish) did find such sentences acceptable. They had apparently overcome negative transfer from Spanish but this acceptance came with a cost; they also found the form acceptable in Spanish and so in becoming more native-like in BP they became less native-like in Spanish. Cabrelli-Amaro et al. proposed that the L3 structure, transferred from Spanish, is reconfigured.

However, a theoretically distinct possibility, and one in line with a more dynamic view of the initial landscape of transfer, is that the English structure (i.e. its combinatorial node in the network) is in fact tagged for L3 use during initial learning. Its presence is undetected at this point in the scalar judgement task because such judgements reflect the overall similarity of Spanish and Portuguese. During L3 immersion, activation of this structure increases through bottom-up processing of BP speech input. In this account, it is the English structure that initiates reconfiguration of the Spanish structure. Evidence of syntactic acceptability is an important point of departure for distinguishing between rival models of transfer, but explicit judgements of similarity are not necessarily sensitive to the possible dynamic variety of the initial transfer landscape. Other converging methods are needed. In the case of S-to-S raising of a dative experiencer in BP, judging the structure as unacceptable during the initial stage of L3 learning might dissociate from a more implicit measure, such as the relative ease of repeating an auditorily presented sentence with that structure when English is the non-source language. If acceptability judgements and sentence repetition dissociate in this way then this would be consistent with the presence of the English structure in the initial syntactic landscape of BP as a L3, despite the transfer dominance of Spanish as the typologically closer language. If so, its activation would strengthen during immersion and increase the activation of the featural configuration that allows S-to-S raising of a dative experiencer in BP and, in line with Cabrelli-Amaro et al. (2015), reconfigure the Spanish structure in the process.

Adapting the control structure for language selection in the third language

Changing representations in the language network is a necessary but not a sufficient condition for L3 speech production. L3 representations (syntactic forms and lexical items) need to be selected for speech production. We start by considering the control processes required for speech production in bilingual speakers. Such control processes must be adapted to allow selection for the L3 but the precise nature of the changes required will depend on how speakers coordinated the use of their two prior languages.

Trilingual language control

Producing a third language, given activation of the two other languages, requires interplay amongst the schemas (Figure 3 illustrates two alternatives). Consider a possible control structure in the bilingual case: one schema is tagged as the mother tongue (Lx) and the other as a “foreign” language (Ly) schema. Such a control structure suffices when there are only two languages competing to select lexical items and constructions for speech output: activation of one reciprocally inhibits the other. However, when there are three languages competing for selection, the foreign language schema will need to differentiate further so that a schema for the L3 can be selected independently of that for Ly – otherwise there will be competition from the previous L2 and difficulty in producing utterances in the L3 at all.

OPEN IN VIEWER

Figure 3.

Configuring the language task schemas in the context of prior language use.

The contexts of language use affect how control processes adapt to the demands of that context (Green & Abutalebi, 2013). Where bilinguals regularly switch between languages in addressing different conversational partners they are arguably more skilled in the competitive control of languages. In contrast, bilingual speakers who primarily code-switch between their languages, especially those engaging in dense code-switching, may be less efficient in using one language to the exclusion of the other. Plausibly, learning a L3 will be affected by prior adaptive changes. Other individual differences aside, between language interference, revealed in overt intrusion errors, may be more common for L3 speakers who predominantly code-switched between their two prior languages compared to those who did not.

In bilingual speakers, individual differences in cognitive control assessed using various neuropsychological tests (e.g. the recursive Tower of Hanoi puzzle that requires individuals to move a stack of discs, ordered by size, one disc at a time from one location to another according to certain rules) are known to predict cross-language intrusions in a picture naming task and inappropriate switching in conversation (Festman, Rodriguez-Fornells, & Münte, 2010). Individual differences in cognitive control are also likely to be predictive of performance errors in the L3. Beginning L3 speakers do report difficulties in retrieving L3 words (Festman, 2006). ³ Surprisingly, individual differences in cognitive control may be most apparent when speakers use a transferred structure, because such a structure has previously been associated with lexical items from the source language and so its use potentially increases lexical competition.

Review and future directions

This paper explored the sources of variability in the trajectories to L3 communicative proficiency in proficient, adult bilingual speakers. We review the main arguments and proposals and then identify key individual differences in language control that are the sources of variability.

Learning a L3 imposes demands on processes concerned with language control, both in terms of representing new information and using that information (Figure 1 illustrates the processing components discussed). Transfer is likely to play an important role in the initial stage of L3 learning and so we characterised the language control processes that sustain transfer. Our characterisation examined the demands implied in three prominent models of syntactic transfer (the TPM, the L2 status factor model and the cumulative enhancement model). We located each model in a two-dimensional control space where one axis indicates the degree of top-down control and the other axis indicates the degree of bottom-up control. Both top-down and bottom-up control processes are necessary for L3 learning. For instance, the former are required if speakers are to learn from formal instruction and the latter are essential if speakers are to induce novel structures from spoken L3 input. These two types of control process are also critical to transfer, but models of transfer emphasise one dimension of control to the exclusion of the other.

A critical claim of the L2 status factor model and the TPM is that a single language determines initial transfer. In assessing such a claim we distinguished between the factors that shape or warp the L3 landscape and the factors that affect which structures actually populate it. What counts as structures within a source language that are available for transfer may vary between individuals as a function of the influence of the L1 and L2 on one another (the permeability of their syntactic forms) and the relative use of structures in the conversational practice of speakers. However such variability does not undermine the critical claim. The critical claim that only one language affects the shape of the landscape of initial transfer requires one boundary condition to be true, namely that the non-source language plays no role in forming the landscape. The TPM and L2 status factor model therefore presume a high degree of top-down control in which the non-source language is completely blocked from shaping and populating the landscape of transfer. In network terms, this means that no structures exclusive to the non-source language are tagged for L3 use despite the high probability that both languages are active and potentially competing to account for novel L3 language input. Effectively, the non-source language is inhibited from accounting for L3 speech input or controlling speech output in the L3. By contrast, the cumulative enhancement model implicates a bottom-up control process in which pre-existing syntactic forms in either of the speaker’s prior languages may capture syntactic forms in the L3 input.

Individual differences in implementing these control processes will yield different starting points for trajectories to L3 proficiency. So, for instance, speakers may differ in their capacity to restrict activation of structures from the non-source language undermining the predictions of the TPM or L2 status factor model regarding initial transfer. Conversely, bottom-up control processes may fail to access all potentially facilitative structures as required by the cumulative enhancement model because structures in the L1 and in the L2 will vary in their relative accessibility as a function of usage. The landscape of transfer for the L3, we suggest, will reflect the weighted outcome of both top-down and bottom-up control processes and so be much more varied than envisaged in current models of transfer.

Transfer also requires change in the language network and in discussing this question we extended the bilingual language network proposals of Hartsuiker et al. (2004). In their network, a syntactic form, such as a passive construction, is depicted by combinatorial nodes. In a bilingual network, when a structure exists in each of the speaker’s prior languages, the particular combinatorial node is linked to two language nodes – one for the L1 and one for the L2. In the case of transfer, then, such nodes can be additionally linked to a L3 language node, that is, tagged as useable in the L3. The TPM and the L2 status factor model envisage all-or-none transfer and so the initial landscape for the L3 will contain syntactic structures that are not part of the L3 grammar. How do speakers overcome negative transfer? Clearly, overcoming negative transfer should not disrupt use of transferred structures that are part of the L3 grammar. We proposed then that initial transfer involves the transfer of dissociable syntactic forms.

Dissociability of syntactic forms allows a modification of part of the network without disruption of the whole. How might modification work? Modifying the network might arise by virtue of explicit feedback and involve top-down control processes that remove the link between the L3 language node and the combinatorial node for the incorrect structure. Once the link is removed, the structure is no longer available for L3 use. Alternatively, the initial link might not be removed. Instead, any activation of that combinatorial node in the context of the L3 would be inhibited. If so, its presence might be unmasked in proficient L3 speakers under conditions (such as stress perhaps) that limit such inhibition. Overcoming negative transfer might also arise over time through bottom-up control processes. Listening to speech input can prime structures that already populate the transfer landscape or lead to the induction of novel structures. Through speaker–addressee alignment (Pickering & Garrod, 2013) such structures become more accessible for use at the expense of non-facilitatory forms that are not part of the repertoire of native L3 speakers. Absence of use does not preclude their continued presence in the language network. Experimental work is needed to distinguish between the alternative ways to overcome negative transfer.

It is insufficient just to change the language network. In order to converse speakers need to be able to select syntactic forms and lexical items in the L3 (rather than those of the L1 or L2) and use the L3 rather than the L1 or L2 as the conversational context demands. Flexibility of language control is afforded by controlling the potential outputs of the language network. Such control processes include ones that inhibit activated but non-target language lemmas and respond to change in the language being used. The coordination of control processes is captured by the notion of language task schemas. A schema acts top-down to control the mapping of a communicative intention (e.g. to speak in the L1) to a speech act by coordinating a set of control processes. For a given speaker, patterns of conversational exchange recur over time and so a schema is effectively a habit of language control. Such schemas can be coordinated either competitively or cooperatively (Green & Abutalebi, 2013). In a conversational context where one language is in use, the relationship is competitive: selection of the schema for the intended language inhibits selection of the other and allows items for that language to be produced overtly but gates items from the non-target language. Where parties to a conversation switch between languages within a conversational turn, there is a cooperative relationship where items from either language can contribute to the incremental production of the utterance (Green & Li, 2014). Bilingual speakers learning a L3 must alter the patterns of control. If bilingual speakers distinguished between their L1 and a non-native language then further differentiation is needed so that, for example, both the schema for the L1 and that for the L2 are inhibited when speaking in the L3. Experimental work is needed to explore individual variety in the interplay between different task schemas in trilingual speakers as a function of their previous pattern of language use.

We focussed on the control and representation of syntactic structures. Utterances clearly combine syntactic forms and lexical material. We understand very little about the mental and neural processes involved in binding different lexical content to common syntactic forms, although such binding is likely to involve cortical and subcortical regions (see Kriete, Noelle, Cohen, & O’Reilly, 2013). Transfer processes will engage the circuits involved and impose demands on them. From a network point of view the use of a transferred syntactic form is economic as it requires creating a link between a pre-existing combinatorial node to a distinct L3 node rather than creating a new combinatorial node as well as a new link to the L3 node. However, the use of a pre-existing structure carries a potential cost: its use has previously been associated with lexical items from the source language. Production of an utterance in the L3 therefore requires overcoming lexical competition to fill the slots in the common syntactic frame. Whether such competition is reduced when an utterance involves a structure unique to the L3 is open to test.

Conclusion

Exploring variability in the trajectories to L3 proficiency, including assessing changes in the processing of the L1 and L2, provides a way to enrich our models of L3 learning. Fundamental to this endeavour is the need to understand the processes of language control. We have argued that theoretical models of transfer differ in the demands they place on top-down and bottom-up processes of control. However, L3 learning and use requires both types of process. Indeed, individual differences in cognitive control, differences in the use of the prior language and the contexts of L3 acquisition ensure variability even in the initial landscape of transfer. Longitudinal approaches combined with a range of methods are required to track variability. Practically, understanding the nature of variability may aid in the creation of learning programmes in the L3 tailored to the individual.