On each training trial, the network was presented with an input (object plus linguistic context), generated as just described, and an appropriate target on the output. The weights in the network, other than those feeding output units for which no targets were available, were then adjusted according to the back-propagation algorithm.
Following each presentation of 1000 input patterns the network was tested on 500 novel inputs generated in the same fashion as the training patterns. There are several options for evaluating the network's performance. We chose to look only at the output unit with the highest activation, unless this unit's activation was not above the response threshold, in which case the network was viewed as not making any overt response at all. Our assumption was that production processes not modeled in our network would force the system to select one word over all of the candidates which might be activated. Thus only the most highly activated output unit was relevant. For each test input, following activation of the network it was determined whether the output unit with the highest activation was above the response threshold and whether that unit corresponded to the appropriate word. Performance for each category of word was measured as the proportion of test trials for which this was true.