Prototype-specific learning for children's vocabulary

Abstract

Several studies suggested that knowledge about the relationship between vocabulary and perceptual objects work as a constraint to enable children to generalize novel words quickly. Children's bias in novel word generalization is considered to reflect their prior knowledge and is investigated in various contexts. In particular, children have a bias to attend to shape similarity of solid objects and material similarity of nonsolid substance in novel word acquisition (Imai and Gentner, 1997). A few studies reported that a model based on Boltzmann machine could explain categorization bias among shape, material and solidity by learning an artificial vocabulary environment (Colunga and Smith, 2000 and Samuelson, 2002). The model has few constraints within its internal structure, but bias emerges through learning artificial vocabulary using simple statistical property about entities' shape, solidity and count/mass syntactical class (Samuelson and Smith, 1999). We proposed a model (prototype-specific attention learning; PSAL) that could learn optimal feature attention for specific prototype of vocabulary. The Boltzmann machine model learns vocabulary in uniform feature space. On the other hand, PSAL learns it in feature space with different metric specific to proximal prototypes. Real children show categorization bias robustly in various learning environment, thus a model should have robustness to various environments. Therefore, we investigated how the two models behave in a few typical vocabulary environments and discuss how prototype-specific learning influence categorization bias