Recursive vocal pattern learning and generalization in starlings.

摘要

Among known communication systems, human language alone exhibits open-ended productivity of meaning. Interest in the psychological mechanisms supporting this ability, and their evolutionary origins, has resurged following the suggestion that the only uniquely human ability underlying language is a mechanism of recursion. This Unique Recursion hypothesis has its roots in the field of generative grammar founded by Chomsky, which has been highly influential in shaping hypotheses about linguistic processes in psychology. The present research follows up on an initial demonstration that a songbird species, the European starling (Sturnus vulgaris), can learn to recognize patterns following a recursive rule (Gentner, Fenn, Margoliash, and Nusbaum, 2006). In a series of three experiments utilizing a context-free (AnBn) and a finite-state (AB n) grammar, I extend the original results to show that many of the assumptions underlying the Unique Recursion hypothesis, including the premise of a monolithic recursive capacity itself, are not easy to reconcile with the learning results obtained for these artificial grammars. Instead of being an automatic, all-or-none ability, recursive pattern-learning depends on constituent fluency, training set, and feedback. Whereas the data are consistent with the context-free grammar making additional demands for subjects, recursion is supported by many interacting learning mechanisms, rather than a qualitatively new mechanism enabling limitless recursion in a single stroke as envisioned in the Unique Recursion hypothesis.