More learnable than thou? Testing metrical phonology representations with child-directed speech

One way to describe a language’s grammar is as a compact system that encodes the regularities of the language. This system allows someone to immediately comprehend and generate novel linguistic items that follow those encoded regularities, and so grammars are often viewed as generative systems. Notably, because languages vary with respect to the specific regularities they have, the generative system can be instantiated in various ways, based on language-specific input (e.g., as a specific set of parameter values in a parametric system or a specific ordering of constraints in a constraint-ranking system). The variables that can be used in a language’s grammar (e.g., the specific parameters or constraints) are defined by the knowledge representation (KR) and so a KR defines the set of possible grammars underlying human languages, based on those variables. The utility of KRs for language acquisition then becomes apparent: if the child already has access to the KR, the hypothesis space of grammars that could encode the regularities of the language is already defined. So, the child already knows which variables in the linguistic environment matter, and so can focus her attention on simply selecting the appropriate instantiation of the KR (i.e., the language-specific grammar), based on those relevant variables. The language acquisition task is about choosing the correct grammar for the language from those defined by the KR. These two aspects of KRs lead to two natural criteria for any KR. The first criterion is the cross-linguistic variation criterion, and states that the right KR should be able to explain the constrained cross-linguistic variation we observe in the world’s languages. The cognitive premise of this kind of argument is that it is surprising to see such limited variation if there is no common underlying KR that humans are drawing their language-specific grammars from. KR theorizing then focuses on identifying the most compact representation than can account for the observed, limited variation. In this vein, Hayes (1995:55) notes, for example, that a successful representation of stress knowledge is one that is “capable of describing all the stress systems of the world’s languages” and is “maximally restrictive”. The second criterion is the learnability criterion, and states that if children have access to