To test our hypothesis that the noun advantage in early acquisition derives from the associative structure of the learning task, we used the most common similarity-based learning procedure in the literature --- a three-layer connectionist network trained with back-propagation. Such a general learning device embodies no prior knowledge about differences between nouns and adjectives, and learning is purely associationist and error-driven.
As in several other recent modeling studies [PSMS92,Sch92], we investigate the behavior of a simple connectionist network which is trained to label a set of patterns representing perceptual inputs to the system. The goal in these studies is to show how the facts of lexical development emerge from the interaction between the learning device and the regularities inherent in the input patterns. In our case, the relevant facts concern the relative ease of learning nouns and adjectives, and the regularities in the patterns concern differences in the way noun and adjective categories carve up the space of input dimensions and co-occur with particular linguistic contexts.
The main difference between our network and other simple connectionist models is our use of a modified form of back-propagation. Back-propagation is suitable in that early word learning in children is ``supervised.'' Adults ask children questions about objects (e.g., ``what is that?,'' ``what color is that?'') and they provide feedback (e.g., ``that's not a dog; it's a horse'') [Cal90,Mer87,Sno77,Woo80]. Supervision for categorization tasks such as our word-learning task, as typically realized in connectionist networks, however, is psychologically unlikely. If separate output units represent the different category responses, standard back-propagation changes the connection weights on each learning trial in a way that encourages the correct response and discourages all other potential responses. This is like the parent saying to the child, ``This is a dog, not a plate, not a cat, not an apple, not a house...'' Parents do not do this but instead explicitly reinforce correct answers (``yes, that's a doggy'') and provide negative feedback only when the child explicitly gives the wrong answer (``that's not a doggy; it's a horse'').
This form of back-propagation is also inappropriate in the present case because in the combined task of naming objects and labeling their attributes, possible responses are not just right or wrong. There are kinds and degrees of wrongness. Consider a big, black, wet dog and the question ``what color is it?'' The answers ``dog'' and ``red'' are both wrong. However, it seems unlikely that parents would respond to these errors in the same way. A toddler who answers the question ``what color is it?'' by correctly naming the dog ``dog'' seems likely to hear a parental response of ``yes, it's a dog, a black dog.'' A toddler who answers the same question by saying ``red'' is likely to hear, instead, a parental response of the sort ``it's not red, it's black.''
Accordingly, we modified the back-propagation algorithm to fit these assumptions about the kinds of feedback provided by parents. Briefly, we provided targets only for a limited number of output words, and we distinguished the kinds of incorrect errors by using distinct targets for them. In the next two sections, we provide a detailed description of the network and the learning rule.