Turning novel names into known names: Understanding referent selection and retention in 24-month-old children and neural networks.

摘要

This thesis explores word learning with 2-year-old children and neural networks. It introduces the probabilistic constraint satisfaction perspective and focuses on how children determine the referent of a novel name. This perspective offers theoretical insight into how children turn novel names into known names---that is, how they initially determine the referent of a novel name and how the underlying cognitive processes that support referent selection can ultimately support full word learning.; Six empirical experiments and three neural network simulations were conducted to investigate the underlying cognitive processes involved in referent selection, focusing primarily on competition and probabilistic constraint satisfaction. Children's behavior in several referent selection contexts was explored using a three-alternative forced-choice task with known and novel objects followed by a stringent retention task. Neural network simulations were conducted with the Hebbian Normalized Recurrent Network, a connectionist model that implements probabilistic constraint satisfaction on a momentby-moment time scale and also learns words via associative learning.; Comparisons between the behavior of the children and the networks helped to elucidate the processes of explicit naming, competition, lexical contrast, constraint satisfaction and associative learning. Specifically, both children and networks showed more evidence of novel name retention when ostensive naming was added to the referent selection task. I argue this increase occurred because ostensive naming simultaneously boosts the novel name-object association and dampens any spurious associations. When auditory competition was added to the referent selection task, both children and networks performed better on the novel than known name trials. Finally, networks predicted that adding lexical contrast to the referent selection would lead to similar results as in the first experiment. However, children behaved as in the competitive condition, suggesting that, at least for 2-year-old children in this task, lexical contrast creates more competition than it helps to alleviate.; Taken together, these results indicate critical roles for competition, constraint satisfaction and associative learning for both referent selection and word learning. Both the empirical and computational research supports the probabilistic constraint satisfaction perspective as an approach to better understanding referent selection and word learning more generally.