The Error-driven Ranking Model of the Early Stage of the Acquisition of Phonotactics: an Initial Result on Restrictiveness *

Nine-month-old infants are already sensitive to the distinction between licit and illicit forms (Jusczyk et al. 1993). They thus display knowledge of the target adult phonotactics at an early stage when morphology is plausibly still lagging behind (Hayes 2004) and the acquisition of the native language lexicon has barely begun (Fenson et al. 1994). How can this early stage of the acquisition of phonotactics be modeled? According to the error-driven learning model, the child is trained on a stream of licit forms, starts from a grammar which corresponds to the most restrictive phonotactics, and slightly relaxes its current grammar whenever it fails on the current piece of data. The model does not require a lexicon, as the slight updates of the current grammar are based on a single piece of data at the time. Furthermore, the model is trained only on licit forms and thus does not require alternations, which might still be unavailable at a stage when morphology is lagging behind. Finally, the model predicts a sequence of grammars that can be matched with child acquisition paths, thus modeling the child’s acquisition gradualness. Section 1 describes an implementation of the error-driven model within Optimality Theory (OT; Prince and Smolensky 2004). Being trained on licit forms, it is easy to guarantee that the phonotactics learned by the model is consistent, namely that it correctly classifies as licit any form which is indeed licit according to the target phonotactics. Yet, the phonotactics learned by the model could fail at restrictiveness: it could incorrectly classify as licit also forms which are instead illicit according to the target phonotactics. Restrictiveness is indeed the main issue of the computational theory of the error-driven model of the early stage of the acquisition of phonotactics, as discussed in section 2. Prince and Tesar (2004) and Hayes (2004) have identified the challenge raised by restrictiveness in the problem of learning a restrictive relative ranking of the faithfulness constraints. Yet, we intuitively expect the relative ranking of the faithfulness constraints to be crucial for the phonology (namely for the specific way in which illicit forms are repaired), but less crucial for phonotactics (namely for the divide between licit and illicit forms). Section 3 formalizes the intuitive condition that the relative ranking of the faithfulness constraints does not matter to describe a certain phonotactic pattern. These con-