MODACQUPHON · Modeling the acquisition of phonotactics

Children learn their target language early, fast and efficiently. For instance, nine-month-old infants already display knowledge of the phonotactics of their target language, namely have been shown to react differently to licit versus illicit sound combinations. Children must thus rely on a remarkably efficient phonotactic learning procedure. What does it look like? According to the error-driven learning model, the learner maintains a current hypothesis of the target adult phonotactics and keeps slightly updating its current hypothesis whenever it makes a mistake on the incoming stream of data from the adult language, until it makes no more mistakes. This learning model is very popular in the current acquisition literature because of its cognitive plausibility: it models the observed acquisition gradualness, as it describes a stepwise progression towards the target adult grammar; it relies on surface phonology, without requiring any knowledge of morphology (that plausibly develops later than phonotactics); and it does not impose unrealistic memory requirements, as it only looks at a piece of data at the time without keeping track of previously seen data. Yet, computational phonology has failed so far to develop a computationally sound implementation of the error-driven learning model. This project aims at filling this gap. Two complementary directions are pursued. An analytical direction geared towards learnability uses tools from Machine Learning to investigate the computational efficiency of the error-driven learning model, focusing on issues such as convergence, correctness, error bounds, and robustness. This analytical strategy is complemented by large scale simulations, that test the model on a large database of infant-directed speech and child acquisition data. Combining a computational perspective focused on learnability with a modeling perspective based on acquisition data will allow my project to break new ground in child language acquisition.