The Computational Problem of Natural Language Acquisition

The paper review work-in-progress on language acquisition in children and robots using combinatory categorial grammar (CCG), building on work by Siskind, Villavicencio, and Zettlemoyer, among others.

CCG is a theory of grammar in which all language-specific grammatical information resides in the lexicon. A small universal set of strictly type-driven, non-structure dependent, syntactic rules (based on Curry's combinators B, S, and T) then "projects" lexical items into sentence-meaning pairs. The task that faces the child in the earliest stages of language acquisition can therefore be seen as learning a lexicon on the basis of exposure to (probably ambiguous, possibly somewhat noisy) sentence-meaning pairs, given this universal combinatory "projection principle", and a mapping from semantic types to the set of all universally available lexical syntactic types.

The paper argues that a very simple statistical model allows children to arrive at a target lexicon without navigation of subset principles, or attention to any attendant notion of trigger other than the notion "reasonably short sentence in a reasonably understandable situation". The model explains the pattern of errors that have been found in elicitation experiments. The linguistic notion of "parameter" appears to be redundant to this process.

The paper goes on to consider some more general implications of the theory, including its application to the phenomenon of "syntactic bootstrapping," and the question of the prelinguistic origin of the combinatory projection principle itself.