Commentary to ‘Multiple Grammars and second language representation’,by Luiz Amaral and Tom Roeper

In the field of second language acquisition, developmental optionality or variability is seen as the defining characteristic: ‘We must consider that perhaps the salient characteristic of interlanguages is that they are … somehow incomplete and in a state of flux’ (Adjémian, 1976: 308). Optionality is the angels and demons of language acquisition. That is, like Escher’s ‘Angels and demons’ drawing, optionality data presents an optical illusion capable of inducing diametrically opposed interpretations in the observers. Statistical learning approaches, such as found in MacWhinney (2005) or Ellis (1996), can correctly model gradualness but fail to make predictions about the basic architecture of the stages or the specificity of transfer, as they deny the existence of grammatical representations underlying linguistic behavior. Generative approaches focus on the formal nature of linguistic representations, and adopted either rule-based or parameter-based systems that translate into predictions about discrete changes in the grammar. The payoff of the principle and parameters model of acquisition coalesces primarily on the side of principles: children obey formal constraints on complex representations at astonishingly young ages, and adult learners have implicit, untaught knowledge of many subtle and obscure properties of a second language (L2) grammar. Parameters were extremely useful in describing the representations in children’s intermediate grammars; there is less agreement as to whether or not this is the case for L2 acquisition (Meisel, 2011). Parameters were of little use in describing the course of development. Most development happens gradually (Paradis and Genesee, 1997), with only selected aspects of learning showing contingent propagation across properties predicted by the model (for some examples, see Snyder, 2007). Roeper (1999) approached the problem of how parametric optionality can exist within a given language with his proposal that speakers are universally bilingual, in the sense that their language competence may be the expression of multiple internal sub-grammars. Each sub-grammar instantiates the establishment of a parameter, the scope of which may be general or limited to classes of lexical items or specific features. Each subgrammar is consistent internally, but the language itself need not be. This approach offers a direct solution to the problem of gradualness: consistent with work in diachronic syntax (Kroch, 1989) and learnability (Yang, 2003), grammar acquisition is categorical, but grammar selection need not be. Amaral and Roeper (this issue; henceforth A&R) take the natural next step for the idea of multiple grammars: to consider the implications of the multiple grammar approach for second language acquisition. They argue that multiple grammars are a natural consequence of the assumptions of minimalism:

(1) The minimalist learner:

 a. assumes no optionality;

 b. faces variability in input;

 c. concludes with multiple rules/subgrammars.

Each rule is a simple statement: a minimal description of a structure. The logical consequences are simple but profound. Parameters cease to function as a description of the learning process (as in ‘parameter setting’), and simply serve to delimit the learning space. Parameters transform the otherwise ill-defined problem of determining for which representation a given piece of input is relevant. Crucially, parameters provide a format that allows learners to track the relevance of specific input to a specific domain.

A pure statistical learner is inherently unable make such decisions: a statistical learning hypothesis might propose that children’s errors with interrogative inversion result from declarative word order, but fail to explain why in some languages, interrogative inversion errors are completely absent in L1 acquisition. What counts as input to learning interrogative inversion? All clauses? Interrogative clauses (including wh and yes/no questions)? Only questions with a unique lexical wh-phrase? Findings such as Valian and Casey (2003) favor the lexically narrow option. The intermediate option (i.e. all types of interrogatives) is ruled out by empirical evidence; Spanish rarely applies subject inversion in yes/no questions, but children learning Spanish show error free acquisition of interrogative inversions with wh-phrases (Pérez-Leroux, 2003). This is the problem of direction of generalization. Nothing inherent in a statistical learner defines what counts as relevant input for learning a given construction. For bilinguals, pure statistical learning predicts an averaging effect across grammars, but cannot explain the systematic structural biases that characterize transfer, such as those found in movement constructions (Strik and Pérez-Leroux, 2011).

Recent proposals in learnability theory conceptualize parameters as pieces of structure (Fodor, 1998; Sakas and Fodor, 2012). A&R take a crucial step further, by requiring that the structural decisions involve the lexical dimension. In their proposal, a minimalist view of L2 learning entails two tasks:

(2) Tasks of the L2 learner:

 a. identification of the set of rules in the target grammar;

 b. defining the lexical scope of rules (lexically narrow or general).

A&R argue in favor of a multiple grammar approach to L2 learning on the basis that:

What would count as evidence for the multiple grammars approach? A&R outline a broad empirical agenda:

The first two points are uncontroversial. The latter deserves some discussion. A&R present an empirical test in a study of pronoun comprehension in Spanish. From a larger group of English learners of Spanish, they selected 27 speakers that approximated native performance, consistently (> 80%) using null pronouns when the embedded subject was co-referent with a matrix subject quantifier.

The resulting subsample of participants, while obeying the Object Pronoun Constraint (OPC) in production, actually showed a preference against it in comprehension. In other words, they rejected a contextually-available specific interpretation for the embedded subject pronoun, choosing instead the coreferent, bound variable reading.

(3) Cada amigo dijo que él iba a traer una bebida.

  ‘Each friend said that he would bring a drink.’ 

A&R suggest that L2 speakers’ possess a competing subgrammar, where coreference between the embedded overt subject pronoun and the matrix subject quantifier is allowed. This intriguing result adds to a growing body of work on comprehension/production asymmetries, which include binding theory. De Villiers et al. (2006) showed that children who fail Principle B by giving an object pronoun a coreferent interpretation to the local subject would not make the parallel comprehension error. Hendriks and Spenader (2006) account for children’s binding asymmetries in terms of bidirectional optimization. This situates comprehension/production asymmetries squarely in the domain of reference, linking their developmental exponent to changes in children’s ability to engage in perspective-taking. Adopting an optimality approach, they show that different sets of constraints operate in production and in comprehension. They propose that children, with their more limited ability to calculate how someone else’s perspective will influence their production choices, fail to restrict local coreference in pronouns.

A&R’s proposal is wider in scope than bidirectional optimization: multiple grammars will be available to speakers whether adults or children, monolingual or multilingual. I would like to bring up two points of evidence to bear into this broader claim. One will remain anecdotal; the second is on record. In 1995, my colleague Bill Glass and I obtained similar results to A&R’s test of Montalbetti’s OPC, with Spanish native speakers. Because we saw these as negative results, and had no explanation on hand, we did not attempt publication. At the time, all we could think was that something must had gone wrong with the method, since we could not tap into a constraint that so clearly revealed itself in production. A&R’s results suggest that OPC comprehension is not random. Their L2 speakers are not performing at chance, but actively opting for the variable interpretation.

The second point pertains to the MG prediction that comprehension will always be broader than production in terms of grammar selection. The most extreme case of this is that of fully receptive bilinguals. Sherkina-Lieber (2011) and Sherkina et al. (2011) report data from speakers of Inuttitut (the Labrador dialect of Inuktitut). These asymmetric bilinguals, while demonstrating incomplete knowledge of their heritage language, retain substantial comprehension abilities and a wide range of grammatical intuitions about morpheme order in this agglutinative language. In production, they have lost all abilities, performing poorly even in simple repetition tasks. One obvious answer to the dilemma presented by these speakers is to attribute comprehension/production asymmetries to processing, or performance, maintaining the classical dichotomy. Alternatively, we could describe receptive bilinguals as having comprehension grammars in the absence of production grammars. A&R skip the obvious, and reach instead for the consistent approach: ‘Processing accounts are usually inexplicit in their relation to representations, but since representations must be involved, it may well be the representation that holds the explanation.’ The proposal that A&R offer is simple, compelling, and difficult. Linguistic theory should model grammar in such a way that makes it possible to account for speakers; even if it involves letting go of language.