Figure 
shows the mean correct responses over
10 separate runs of the network. As can be seen, outputs referring to
Small categories are learned faster than the ones referring to Large
categories (
). The difference is smaller than in Experiment 1 probably
because the ratio of Large-to-Small volume is smaller: 16 to 1 in
this experiment, but 216 to 1 in Experiment 1.
: Experiment 2: Category Volume.
Performance is the proportion of test items for which the highest
overt response was correct.
Responses are averaged over 10 separate runs of the network.
The network also readily learned the association between one linguistic context input and the class of Large-volume outputs and between the other linguistic input and the class of Small-volume outputs. As in Experiment 1, we examined ``within-part-of-speech errors;'' here the Small-volume and Large-volume categories represented the two parts of speech. At the start of learning, ``within-part-of-speech'' errors comprised (as expected by chance) about half the errors for both Small-volume and Large- volume targets (.50 of the errors given a Small-volume targets and .45 of the errors given a Large-volume target). After 1000 training trials, however, within part-of-speech errors predominated, .88 of all errors given a Large-volume target and .81 of the errors given a Small-volume target. These results again demonstrate the role of word-word associations in the network's learning.
In sum, this experiment shows that differences in the volume of a category, one of the differences that exists between common nouns and dimensional adjectives is sufficient to create an advantage in learning. This is not an unexpected result, given all that is known about the importance of within-category similarity to similarity-based learning. But it is a result that is consistent with the idea that developmental differences between the early acquisition of nouns and adjectives could derive from processes no more complex than those embodied by a three-layer connectionist network.
