Goldrick, Putnam and Schwarz (henceforth Goldrick et al.) base their computational account of code mixing on two premises, (1) that code mixing is subject to grammatical principles and (2) that “mental representations in bilingual speakers incorporate blends of structures from each language”. While there is copious evidence for the first premise, the second is far less well empirically supported. The authors imply that the second premise, the assumption of blend representations, is supported by Kroll & Gollan's (Reference Kroll, Gollan, Goldrick, Ferreira and Miozzo2014) review of speech planning in two languages, but this is not the case. Kroll & Gollan (Reference Kroll, Gollan, Goldrick, Ferreira and Miozzo2014:1) do indeed discuss evidence for the “parallel activation of the bilingual's two languages” but in terms of competition between alternatives rather than blended representations. The notion of blend is of course familiar to code-switching researchers from work on bimodal bilingual code mixing, as Goldrick et al. acknowledge, but here the term refers to “simultaneously produced English words and ASL signs” (Emmorey, Borinstein, Thompson & Gollan, Reference Emmorey, Borinstein, Thompson and Gollan2008: 43). In speech the term has long been used to refer to a phonological mixture of two words, often in error: Fromkin (Reference Fromkin1971: 40) gives the example of clarinola as a blend of clarinet and viola. Poulisse (Reference Poulisse1999:172) finds blends between L1 Dutch and L2 English lexical items to be relatively rare, and argues that morphological and syntactic coding are language-specific.

There is thus no evidence in the literature for a consensus on the existence of blend representations in code-switching. On the contrary, there is much more empirical evidence for an aversion to doubling by bilinguals. Many code-switching corpora (unfortunately not examined by Goldrick et al.) exhibit asymmetry between the two languages in contact, with one language providing the morphosyntactic frame and the other the source of insertions of words or phrases into that frame (cf. Deuchar, Reference Deuchar2006, Deuchar, Muysken & Sung-Lan Wang, Reference Deuchar, Muysken and Wang2007, Muysken, Reference Muysken2000, Myers-Scotton, Reference Myers-Scotton2002). A computational account of this widespread phenomenon, which seems to favour competition rather than blending, could have been very enlightening.

Instead, Goldrick et al. focus on what they themselves, following numerous other researchers, acknowledge to be the “exceedingly rare” (p.14) phenomenon of doubling. An important methodological issue which they do not address is how to treat infrequent phenomena in corpora, and whether doubling could be a ‘performance error’ given its rarity. Myers-Scotton (Reference Myers-Scotton1993) presents a detailed account of how doubling may arise, assuming Levelt's (Reference Levelt1989) Speaking model and concludes that “double morphology may result from misfiring at some point in production” (Myers-Scotton, Reference Myers-Scotton1993:135). A similar view is apparently adopted by Muysken (Reference Muysken2000:105) who suggests that doubling involves an “adjustment in the planning of the sentence” and a “syntactic break”.

Far from acknowledging these previous suggestions that doubling may represent error, Goldrick et al. argue that the occurrence of doubling is “constrained by grammatical principles” and involves “equivalence for grammatical features” (p. 16). In contrast, though, they argue that doubling involves violation of “a preference for congruent grammatical patterns in code-mixing” (p. 15). Appealing to a strongly dispreferred structure-type as the basis for a (very powerful Gradient Symbolic Computation-oriented) model of bilingual competence already raises questions, but an additional problem arises from the fact that the single Tamil–English example at the heart of their proposal is not representative of the doubling structures that Goldrick et al. themselves present (pp.13–14), let alone those that have been discussed in the literature.

While Goldrick et al. state that there is a preference for the doubling of functional elements, their central Tamil–English example involves the doubling of lexical verbs. And while they may be able to claim “equivalence for grammatical features” in this case, it cannot be claimed cross-linguistically for all functional elements. Since Rizzi's (Reference Rizzi and Haegeman1997) paper on the “fine structure” of CP, it has become well known that the structure of functional heads such as complementisers exhibits cross-linguistic variation. For example, Goldrick et al.’s example (9) includes ‘doubling’ of the English complementiser that and the Korean complementiser ko. However, Bhatt & Yoon (Reference Bhatt, Yoon and Bates1992: 49) point out that these are not equivalent since they behave differently in indirect questions (see also Kim, Reference Kim, McClure and den Dikken2011). Similarly, Goldrick et al. use their example (3) to demonstrate ‘doubling’ of complements in English and Japanese despite the demonstration by Kuwabara (Reference Kuwabara2013) that complementiser structure is very different in Japanese compared with English. Thus although the term doubling implies the cross-linguistic grammatical equivalence that is claimed by Goldrick et al., recent work suggests that such equivalence at the functional level is an illusion. In fact, this non-equivalence may explain the appearance of the phenomenon in the first place.

An alternative to treating doubling as performance errors, then, would be to treat them as indicative of possible convergence. Convergence (see e.g., Backus, Reference Backus2005; Silva-Corvalán, Reference Silva-Corvalán1986; Toribio, Reference Toribio2004) is a phenomenon well worth modelling in a way that encompasses its other manifestations, including omission. Goldrick et al.’s faithfulness constraint penalises omission of constituents considerably, and yet this is a phenomenon which is associated with grammatical convergence in language contact. Davies & Deuchar (Reference Davies and Deuchar2014), for example, document the frequent omission of utterance-initial auxiliaries in Welsh, resulting in SVO structures which are more congruent with English, the language with which Welsh is continually in contact. A model representing the widespread phenomenon of grammatical convergence would also have the advantage of helping to account for contact-induced language change.