Magri, G.
CNRS, University of Paris 8
Nine-month-old infants are already sensitive to the distinction between licit and illicit forms drawn by the target adult phonotactics (Jusczyk et al. 1993). Yet, they are still blind to alternations, as morphology still lags behind (Hayes 2014). How can they manage to acquire phonotactics from licit forms only? According to the error-driven phonotactic learning model, the child is trained on a stream of licit forms, starts from the most restrictive initial phonotactics, and slightly relaxes it whenever it fails on the current piece of data. The model does not require a lexicon (which might still be unavailable at this age) and predicts a sequence of grammars that can be matched with child acquisition paths (Gnanadesikan 2004). Being trained on licit forms, it is easy to guarantee that the phonotactics learned by the model is consistent: it correctly rules forms which are indeed licit according to the target phonotactics. Yet, the learned phonotactics could be unrestrictive: it could rule in also forms which are instead illicit according to the target phonotactics (Prince and Tesar 2004). In this talk, I tackle the problem of restrictiveness of error-driven learning within the framework of Optimality Theory (OT; Prince and Smolensky 2001). Knowledge of phonotactics is modeled in OT through a ranking of markedness and faithfulness constraints. I show that the relative ranking of the faithfulness constraints turns out to be irrelevant for phonotactics in the vast majority of cases. I then develop an implementation of the error-driven phonotactic learning model within OT and prove that it is restrictive in this vast majority of cases. These results illustrate an approach to cognitive modeling based on the systematic investigation of the algorithmic implications (e.g., implications for restrictiveness) of peculiar typological properties (e.g., the fact that the relative ranking of the faithfulness constraints does not matter).