The nouns we sought to model are the common names for concrete objects, such nouns as bottle, cup, mom, dog, bed, and cookie. The idea that the extensions of such nouns are relatively small, compactly shaped, and non-overlapping has been offered repeatedly in the literature (e.g., Rosch, 1973; Markman, 1989, Medin & Ortony, 1989). Further, [Ros73b] has reported empirical evidence in support of these claims and Mervis (1987) has presented evidence that when mature usage does not fit this characterization, adults in their speech to children shift their use of nouns to keep the extensions compactly shaped and non-overlapping.
However, there are other findings in the literature that might be interpreted as showing that common names are not compact but are, rather, adjective-like in their emphasis of a single dimension. These findings concern the so-called shape bias in early noun learning (see Smith, 1995 for a review).
Specifically, in novel word learning tasks, when a novel rigid object is named by a count noun, young children systematically generalize the newly learned name to novel objects by their shape ignoring such properties as color and texture. This shape bias in early noun learning fits well with Biederman's (1987) and Rosch's (1973)
earlier results showing that adults recognize common objects principally by their shape. Do these results, then, suggest the real nouns learned by young children are not compact but are rather like adjectives in being constrained principally on one dimension. The evidence on the shape bias in early word learning is quite compelling, but we believe the inference from this fact about the non-compactness of noun extensions is wrong.
First, the complete evidence on the shape bias does not suggest exclusive attention to shape in children's early word learning. Rather, children attend to shape when objects are rigid but attend to texture and color when they are nonrigid, and even with rigid objects, children attend to shape and texture when the objects have eyes (see Smith, 1995 for a review). Thus, the total pattern suggests that while young children often emphasize shape in their early word generalizations, it is not to the complete exclusion of other properties. Rather, children attend to other properties and shift attention weights as a function of those properties. Second, although shape may be important to determining membership in a specific category, for real categories (as opposed to those used in artificial word learning tasks), other properties are also clearly predictive of category membership. Thus dogs do not just have a characteristic shape, they have characteristic colors, surface properties, and manners of movement. Thus, the extensions of the nouns that children encounter are relatively compact. Finally, shape is not a simple dimension but is composed of many sensory dimensions; constraints on the shape of instances will thus make for more compact category extensions than constraints on, for example, wetness or color. In sum, the extensions of the real nouns that children learn early may not be hypercubes in the space of all possible objects, but all that we know indicates that they are much more compact than dimensional adjectives.