Commitment in L3 relationships: Sacred vows or polyamory?

In Westergaard’s (2021) updating of the Linguistic Proximity Model (LPM), she renews her argument for incremental, conservative learning rather than parameter setting and reasserts her vision of parsing as the primary acquisition mechanism, before arriving at the heart of her article, where she mounts a sustained and persuasive defense of property-by-property transfer from either the first language (L1) or second language (L2) in third language (L3) acquisition. While the abandonment of macro-parameters and the argument for acquisition through parsing may be relatively mainstream positions, the idea that the acquisition of an L3 is not based on a single previously acquired language, but that it can draw on any other language pre-existing in the same mind, has engendered considerable debate in the newly established subfield of L3 acquisition research. After briefly considering background assumptions about acquisition in general, I evaluate this new statement of the LPM in the light of early generative research on transfer, recent models of L3 acquisition, and L1 attrition, in an effort to discern the possibility of love triangles between languages in the multilingual mind.

It is now widely accepted in linguistic approaches to L2 acquisition that the promise of Principles and Parameters has ultimately remained unfulfilled (e.g. Lardiere, 2009). Just as previous research had cast doubt on parametric explanations of null subjects, binding, and verb-raising, Westergaard (2021) provides convincing evidence against V2 as a whole-language parameter. In Norwegian, V2 is obligatory when the wh-element is long, discourse-dependent when the wh-element is monosyllabic, and almost unattested when the element ka ‘what’ is sentence-initial (as previously discussed by Westergaard, 2009b). So, while syntactic operations may be widely generalized, and even extend to all relevant structures in a language, acquisition accounts cannot rely on a simple language-wide parametric switch. This kind of structure-by-structure account also seems best able to explain that fact that as languages change over time, there are often remnants of previously instantiated ‘parameters’. Thus, contemporary English still exhibits V2, but only in fronted negative adverbials, including lexical adverbs, e.g. Only in Maine will you find such tasty lobsters; as well as negative PPs, e.g. Not till I got home did I realize that I had fallen in love.

A core aspect of Westergaard’s (2009a) Micro-cues Model is its attempt to explain acquisition entirely in terms of the parser. That the parsing of input is somehow the mechanism that drives the acquisition of syntax is common to many alternative accounts, including the Modular Cognition Framework (MCF; Sharwood Smith and Truscott; 2014); Autonomous Induction (Carroll, 2001); Processability theory (Pienemann, 2005); and Parser as the Language Acquisition Device (PLAD; Dekydtspotter and Renaud, 2014). While parsing is central to each approach, there is considerable variation in ideas of how exactly parsing-to-learn might work, and thus far no approach has furnished empirical evidence persuasive enough to convince the field of its distinct vision. Westergaard’s (2021: 390) idea that ‘parsing is the only mechanism for language learning’ seems to be very much a case of overreach. Several aspects of lexical acquisition, including recovery from over- and underextension of reference, seem to require more than an adapted version of syntactic parsing. Even if parsing were extended to aspects of phonology, this would require a distinct, independent parser, given that syntactic and prosodic boundaries do not match, and seem to be generated by independent representational modules (for examples and discussion, see Jackendoff, 1997: 26–27). Moreover, Westergaard (2021) suggests that micro-cues, unlike Fodor’s (1998) treelets and Lightfoot’s (1999) cues, are not provided by UG but are language-specific. The argument for the need to situate cues in linguistic contexts such as clause type or verb type is well-taken, but this raises the question of whether the mechanism is sufficiently constrained. Can any piece of structure be a micro-cue? How do we define what could serve as a cue?

Another key element of the Micro-cues Model is the assumption of conservative learning. Generalization is understood to proceed in small steps, ‘typically involving the addition of a sub-category, a lexical item, or feature to an already existing micro-cue’ (Westergaard, 2021: 384). The claim that children do not usually overgeneralize, thus reducing the need for unlearning, might seem to fly in the face of common understandings about child language development, as so many studies have focused precisely on overgeneralization. However, even Pinker (2013: 375), in one of the most well-known monographs dealing with L1 overgeneralization, argues that despite constrained instances of creative overapplication of rules, ‘conservatism is the rule’. For similar conclusions for overgeneralization of past-tense -ed, see Marcus et al. (1992), and for an overview of conservative, step-by-step acquisition of inversion in questions, see Ingram (1989: 457–65). In this respect, Westergaard’s (2021) approach seems to be very much in line with other well-grounded studies of L1 acquisition of morphology and syntax.

The debate concerning the nature of L3/Ln acquisition appears to have largely settled around two opposing viewpoints (excluding other hypotheses that have gained less traction). On the one hand, in support of the Typological Primacy Model (TPM), authors such as González Alonso and Rothman (2017), Rothman (2011, 2015), and Schwartz and Sprouse (2021) argue that an L3 learner selects either the L1 or the L2, on the basis of perceived typological similarly to the L3, which is then copied as a single, coherent system to be restructured as the emerging L3. On the other hand, the LPM developed by Mykhaylyk et al. (2015) and Westergaard et al. (2017) and the Scalpel Model of Slabakova (2017) both eschew the idea of wholesale transfer exclusively from a single language, and argue that crosslinguistic influence may be property-by-property from any other language co-existing in the same multilingual space. That L3 acquisition might proceed on the basis of copying an entire L1 or L2 system in advance of attaining even an intermediate level of the L3, before even recognizing whether the L3 has, for example, determiners, or relative pronouns, or topic-marking, would mean that a kind of ghost grammar is brought into being, awaiting embodiment.

This is expressed by Schwartz and Sprouse (2021: 16) as follows: ‘at a rather early point of exposure to PLD [primary linguistic data], L3ers commit themselves to what we like to call the BIG DECISION, i.e. the brain somehow subdoxastically selects the grammar of one of the previously acquired languages as the basis for a new grammar.’ That is, after an initial stage in which the L3 may be attracted to either the L1 or the L2, a ‘sacred vow’ is undertaken such that the L3 enters into a relationship with only one other language, and may no longer mingle with any other languages in the same brain. If the L2 is chosen as a source of transfer, the L3 may no longer consort with the L1. This commitment to transfer is made regardless of future respective language competence. However, the balance between languages often shifts over time – attrition can set in, periodically or drastically. Proficiency can wax and wane (consider summer vacations for students, study abroad, visiting family members, etc.). Understandings of more dynamic relationships of transfer require other relationship metaphors: perhaps polyamory or elective affinities. A property-by-property account of crosslinguistic influence seems a more promising way forward for several reasons.

First, if all languages remain active in the multilingual mind, they can be accessed without extra copies having to be made. As Westergaard (2021) notes, psycholinguistic evidence confirms that all languages in a multilingual brain are active by default during language processing and when one is used the others must be inhibited (Kroll et al., 2012). Schwartz and Sprouse (1996) may have understood Full Transfer as involving an initial, full copy of the L1; however, if the only other language in the mind is the L1, then we cannot tell whether transfer is based on a whole-copy-restructuring model, or whether transfer occurs at different times throughout the acquisition process. As Westergaard (2021) points out, White (2003) was more tentative in her assessment of this issue, and it is also the case that Hawkins (2001: 66) explicitly stated that ‘One possibility is that L1 influence in L2 grammar-building only occurs at the point where the relevant representation in the grammar is being constructed.’ This is thoroughly compatible with Westergaard’s (2021: 389) assertion that ‘The complete grammar of the L1 remains active (so no need to make a copy of it).’ From the perspective of Hawkins (2001) and Westergaard (2021), the influence from L1 relative clause structure, for example, will be non-existent at earlier stages when the interlanguage is smaller than a fully realized language system and contains no relative clause structures at all.

Second, the LPM should be understood as exempt from the criticism of piecemeal transfer in Schwartz and Sprouse (1996), as this quite distinct from the notion of property-by-property transfer in the current discussion (pace Schwartz and Sprouse, 2021). The original argument was in the context of accounts such as Vainikka and Young-Scholten (1994), who argued that L1 functional projections were not available for transfer, or Eubank (1996), who considered L1 functional feature values to be inaccessible. According to the LPM/Scalpel Model, nothing is inaccessible; everything in the pre-existing languages of the mind is potentially available for transfer, so the issue is not one of restrictions on transfer.

Third, the full copy account seems to be inexorably tied to the traditional acquisition contexts examined in typical early L2 research, in which a monolingual native speaker was assumed to have a steady-state L1, and to be acquiring the L2 in serial fashion. However, subsequent research has questioned the very existence of a steady state and has examined other forms of transfer in the bilingual mind, which cannot be so easily explained by a full copy theory. For example, a dominant L2 can exert crosslinguistic influence on the L1, presumably without creating a copy of the L2 to replace the L1. Transfer can be in either direction, depending on factors such as level of activation of each language, and the relative robustness of particular representations determined in part by frequency of use and recency of access (Bardovi-Harlig and Stringer, 2017; Cook, 2003; Schmid and Köpke, 2017; Sorace, 2000). In the area of pronominal binding, Gürel (2002) found that both L1 English acquirers of Turkish in Istanbul and L1 Turkish speakers undergoing attrition in North America had the same error patterns under the influence of dominant English. In such cases, if there is no qualitative difference between L1 effects on the L2 and L2 effects on the L1, it makes more sense to think in terms of crosslinguistic influence (CLI), rather than the metaphor of transfer, as language systems are not cloned and shipped elsewhere. Other acquisition contexts that complicate a full copy hypothesis include: acquisition of L1 social or regional dialects; creole formation with multiple substrates; shifts in dominance over time between L2 and L3; and the kind of functional multilingualism found throughout the postcolonial world, where the target language is in fact a multilingual code repertoire (Stringer, 2015).

Fourth, while the research questions are still in play and new evidence is still forthcoming, there do seem to be a few serious empirical problems for the full copy approach of the TPM. As Westergaard (2021) notes, in their review of L3 acquisition research, Puig-Mayenco et al. (2020) found 17 studies that revealed evidence of property-by-property transfer involving both the L1 and the L2. As for TPM-inspired studies, current formulations appear to be unfalsifiable, as González Alonso and Rothman (2017) allow for piecemeal transfer both before and after the moment of full, wholesale transfer. This gives the TPM wiggle-room to provide alternative accounts of any counterevidence of property-by-property transfer. Citing the need for ‘cognitive economy’ to explain wholesale transfer just for those stages of acquisition after the initial stage but prior to more advanced acquisition falls short of actual evidence against property-by-property transfer throughout the L3 acquisition process.

One thing that emerges from this debate is the vibrancy of current research in this domain. The strong hypotheses put forward by González Alonso and Rothman (2017), Schwartz and Sprouse (2020), and Slabakova (2017) have provided clear points of reference so that future empirical studies can shed light on the nature of the initial state, transitional stages, and ultimate possibilities in multilingual acquisition. In this context, Westergaard (2021) provides a well-crafted and thought-provoking model of acquisition, according to which languages cohabiting in the same mental space may freely interact with each other. This seems to fit well not only with current research on L3 acquisition, but with research on multilingualism across a range of populations.