Distributional learning and language acquisition

One of the central problems for the cognitive sciences has been to account for language acquisition; the difficulty arises in coming up with a theory that simultaneously satisfies three conflicting goals (Jackendoff 2011); accounting for the variety and complexity of attested natural languages; explaining the acquisition of these from the available data, while not assuming too specific an innate endowment, which is implausible from an evolutionary perspective. In this talk we will outline a candidate solution, based on a mathematically precise theory of grammatical inference based on distributional learning techniques. These techniques, only recently developed (e.g. Clark & Eyraud 2007, Clark & Yoshinaka 2014), show how large classes of very powerful grammars can be learned from example sentences by decomposing the sentences into substrings and contexts which correspond respectively to the yields of subderivations and derivation contexts. These methods are applicable to a wide range of grammar formalisms, including mildly context-sensitive grammar formalisms such as Multiple context free grammars, which are equivalent to Stabler's Minimalist Grammars. We will describe the current state of this theory and the (in)compatibility of these models with standard models of syntax, make some precise predictions about the range of possible grammars based on learnability considerations, and discuss how plausible this is as the basis for a theory of language acquisition.