It is well known that children can learn words rapidly. Recently, several studies suggested that knowledge about the relationship between vocabulary and perceptual objects works as a constraint to enable children to generalize novel words quickly. Many experimental results of novel word generalization task supported this hypothesis, but its mechanism remains unclear. In this study, we examined a past proposed model explaining its mechanism and showed that they could not simulate novel word generalization task well in certain conditions. Therefore, in stead of the previous model, we proposed a model that could learn optimal feature attention for specific prototype. Our proposed model works well even with a multidimensional vocabulary set including rich perceptual information that the past model could not work. It suggested that statistical learning could be powerful enough to solve feature selection problem even in noisy information source. Furthermore, it also suggested that the basis of chilren's word learning was prototype-specific feature attention.