Goldrick, Putnam and Schwarz (Reference Goldrick, Putnam and Schwartz2016) propose a very explicit account of code mixing, which nicely draws implications from the well-established findings of coactivation during bilingual language production, and that code mixing is constrained by grammatical principles. This precise characterization will undoubtedly serve as a useful basis for further research on code mixing, second language sentence production, and syntactic learning. However, there are three issues with the account that require further elaboration.
Does having multiple word orders entail flexibility?
When defining the first principle of their account, Goldrick et al. argue that English requires the word order SVO, whereas Dutch allows “flexibility between SVO, SOV, and VSO.” This is incorrect. All three orders indeed occur in Dutch, but they cannot be chosen flexibly. SVO is the order of the main clause and SOV is the order of a subordinate clause starting with “omdat” (because) (1a). But subordinate clauses starting with “want” (as) (2) have SVO order. VSO is required when a unit (e.g., a Wh-word) is moved sentence-initially (see Koster, Reference Koster1975, for a generative analysis of Dutch word order). In contrast, order of auxiliary verb and participle within SOV clauses (1a-b) is flexible. Perhaps such facts of Dutch grammar can be captured by a more elaborate list of constraints. But it does seem that an analysis that simply ranks these word orders according to their overall probability misses an essential point about this variation: surely SVO, VSO, and SOV occur with certain probabilities of x, y, and (1 – (x+y)), but if we know the syntactic and lexical context, we can predict the order with certainty.
-
(1a) Zij belde de politie omdat de zakkenroller haar portemonnee had gestolen
Lit.: She called the police because the pickpocket her wallet had stolen
-
(1b) Zij belde de politie omdat de zakkenroller haar portemonnee gestolen had
Lit.: She called the police because the pickpocket her wallet stolen had
-
(2) Zij belde de politie want de zakkenroller had haar portemonnee gestolen
Lit.: She called the police as the pickpocket had her wallet stolen
“She called the police <because/as> the pickpocket had stolen her wallet”
Is semantics shared or separate across languages?
The account of doubled elements rests on the very strong assumption that there is co-activation, to varying degrees, of two separate semantic representations, one for English gave and one for Tamil kodutaa (see Goldrick et al.’s example 15). This assumption is crucial for the analysis of doubled elements, because compliance with the faithfulness constraint now promotes the rank of the doubled construction so that it acquires some probability. However, the authors do not provide an independent motivation for the assumption of a doubled representation in the semantics. In fact, the more fundamental question can be asked whether semantic representations in L1 and L2 are shared or separate. Note that psycholinguistic theories (e.g., Kroll & Stewart, Reference Kroll and Stewart1994) traditionally assume a common semantic system for both languages or a feature-based system with high overlap between L1 and L2 semantic representations (Van Hell & De Groot, Reference Van Hell and De Groot1998). There is (masked) priming of non-cognate translation equivalents across languages (Schoonbaert, Duyck, Brysbaert & Hartsuiker, Reference Schoonbaert, Duyck, Brysbaert and Hartsuiker2009) suggesting shared or overlapping representations. Relatedly, cross-language structural priming studies showed stronger priming if head verbs in prime and target sentence were translation equivalents (Schoonbaert, Hartsuiker & Pickering, Reference Schoonbaert, Hartsuiker and Pickering2007). In light of such theories and findings, the assumption of discrete and fully separate semantic representations for each language is surprising.
Is the account overparametrized?
The paper refers to a learning algorithm, but does not describe the algorithm, parameters, training data, or fitting procedures in detail. This leaves it unclear how many free parameters there are in the unilingual case. But additionally, the adaptation to bilingualism introduces further parameters that for instance determine the strength of each language, the strength of each of the discrete semantic representations, the strength of the ‘boost’ to a constraint depending on the head word's language, and so on. Furthermore, all terminal and non-terminal nodes have activation values that interact with the contraints. If the goal is to provide a general, descriptive framework that demonstrates how code mixing could be analyzed, such aspects of training and fitting might be viewed as uninteresting details of implementation. But if the more ambitious goal is to develop a precise account that makes testable and quantitative predictions, clarification is needed of what data sets or experiments can be used to test the account and how the number of free parameters compares to the number of to be fitted data points. For now, the seemingly large number of degrees of freedom raises questions about the falsifiability of the account.
Goldrick, Putnam and Schwarz (Reference Goldrick, Putnam and Schwartz2016) propose a very explicit account of code mixing, which nicely draws implications from the well-established findings of coactivation during bilingual language production, and that code mixing is constrained by grammatical principles. This precise characterization will undoubtedly serve as a useful basis for further research on code mixing, second language sentence production, and syntactic learning. However, there are three issues with the account that require further elaboration.
Does having multiple word orders entail flexibility?
When defining the first principle of their account, Goldrick et al. argue that English requires the word order SVO, whereas Dutch allows “flexibility between SVO, SOV, and VSO.” This is incorrect. All three orders indeed occur in Dutch, but they cannot be chosen flexibly. SVO is the order of the main clause and SOV is the order of a subordinate clause starting with “omdat” (because) (1a). But subordinate clauses starting with “want” (as) (2) have SVO order. VSO is required when a unit (e.g., a Wh-word) is moved sentence-initially (see Koster, Reference Koster1975, for a generative analysis of Dutch word order). In contrast, order of auxiliary verb and participle within SOV clauses (1a-b) is flexible. Perhaps such facts of Dutch grammar can be captured by a more elaborate list of constraints. But it does seem that an analysis that simply ranks these word orders according to their overall probability misses an essential point about this variation: surely SVO, VSO, and SOV occur with certain probabilities of x, y, and (1 – (x+y)), but if we know the syntactic and lexical context, we can predict the order with certainty.
(1a) Zij belde de politie omdat de zakkenroller haar portemonnee had gestolen
Lit.: She called the police because the pickpocket her wallet had stolen
(1b) Zij belde de politie omdat de zakkenroller haar portemonnee gestolen had
Lit.: She called the police because the pickpocket her wallet stolen had
(2) Zij belde de politie want de zakkenroller had haar portemonnee gestolen
Lit.: She called the police as the pickpocket had her wallet stolen
“She called the police <because/as> the pickpocket had stolen her wallet”
Is semantics shared or separate across languages?
The account of doubled elements rests on the very strong assumption that there is co-activation, to varying degrees, of two separate semantic representations, one for English gave and one for Tamil kodutaa (see Goldrick et al.’s example 15). This assumption is crucial for the analysis of doubled elements, because compliance with the faithfulness constraint now promotes the rank of the doubled construction so that it acquires some probability. However, the authors do not provide an independent motivation for the assumption of a doubled representation in the semantics. In fact, the more fundamental question can be asked whether semantic representations in L1 and L2 are shared or separate. Note that psycholinguistic theories (e.g., Kroll & Stewart, Reference Kroll and Stewart1994) traditionally assume a common semantic system for both languages or a feature-based system with high overlap between L1 and L2 semantic representations (Van Hell & De Groot, Reference Van Hell and De Groot1998). There is (masked) priming of non-cognate translation equivalents across languages (Schoonbaert, Duyck, Brysbaert & Hartsuiker, Reference Schoonbaert, Duyck, Brysbaert and Hartsuiker2009) suggesting shared or overlapping representations. Relatedly, cross-language structural priming studies showed stronger priming if head verbs in prime and target sentence were translation equivalents (Schoonbaert, Hartsuiker & Pickering, Reference Schoonbaert, Hartsuiker and Pickering2007). In light of such theories and findings, the assumption of discrete and fully separate semantic representations for each language is surprising.
Is the account overparametrized?
The paper refers to a learning algorithm, but does not describe the algorithm, parameters, training data, or fitting procedures in detail. This leaves it unclear how many free parameters there are in the unilingual case. But additionally, the adaptation to bilingualism introduces further parameters that for instance determine the strength of each language, the strength of each of the discrete semantic representations, the strength of the ‘boost’ to a constraint depending on the head word's language, and so on. Furthermore, all terminal and non-terminal nodes have activation values that interact with the contraints. If the goal is to provide a general, descriptive framework that demonstrates how code mixing could be analyzed, such aspects of training and fitting might be viewed as uninteresting details of implementation. But if the more ambitious goal is to develop a precise account that makes testable and quantitative predictions, clarification is needed of what data sets or experiments can be used to test the account and how the number of free parameters compares to the number of to be fitted data points. For now, the seemingly large number of degrees of freedom raises questions about the falsifiability of the account.