Comparing exemplar and prototype models of categorization

Canadian Journal of Experimental Psychology, Sep 1997 by Stephen Dopkins, Theresea Gleason

RESULTS

The results for the learning phase are summarized in Table 2. The table shows the proportions of correct responses for the entire learning phase and for the last 120 trials (the last quarter of the trials in the learning phase). In both cases, performance is broken down as a function of category and as a function of the way the axes of the parameter space were assigned.

Consider first the results for the entire learning phase. An analysis of variance showed that the proportion of correct responses did not differ as a function of category, F(1,30)

In the data for the last quarter of the learning trials, these differences disappeared. An analysis of variance showed that the number of correct responses did not differ as a function of category, F(1,30) = 1.49, MS^sub e^ = .005, p > .05, or axis assignment, F(1,30) = 2.02, MS^sub e^% = .019, p > .05, and that the effects of category and axis assignment did not interact, F(1,30)

The results for the crucial test trials are summarized in Table 3. Proportions of assignment to the two categories are broken down as a function of test stimulus and axis assignment. (Note that the proportions for the two categories do not always add up to 1. This is because participants sometimes emitted responses other than "1" or "2"). Participants assigned the test stimuli to the compact category more frequently than to the dispersed category. A paired t-test showed that the number of assignments to the compact category was significantly greater than the number of assignments to the dispersed category, t(31) = 2.78, SD = 12.89, p

An analysis of variance was conducted to find out whether the response pattern on the test trials differed as a function of test stimulus and axis assignment. The dependent variable in this analysis was number of compact category responses. The analysis showed that the numbers of these responses did not differ as a function of test stimulus, F(2,60) = 1.56, MS^sub e^ = 2.66, p > .05, or axis assignment, F(1,30) = 1.13, MS^sub e^ = 13.99, p > .05, and that the effects of test stimulus and axis assignment did not interact, F(2,60)

The results for the rest of the generalization trials are summarized in Table 4. Participants made more dispersed than compact category responses, F(1,31) = 14.18, MS^sub e^ = 509, p

DISCUSSION

Participants tended to assign the test stimuli to the compact category rather than the dispersed category. We would like to interpret these results as evidence for the exemplar theory and against the prototype theory. This conclusion follows if participants assigned the test stimuli to the compact category because the stimuli were more similar to the exemplars of that category.

However, there may be other explanations for these results. First, the results may reflect a pattern of response bias. Recall that some of the stimuli presented during the generalization phase had also been presented during the learning phase. More of these "old" stimuli belonged to the dispersed category than to the compact category. Presumably, participants tended to assign these "old" stimuli to the dispersed category. It is possible that participants developed a response bias toward the compact category in an attempt to equalize the numbers of responses to the two categories. One argument against this account is that participants, overall, made more dispersed than compact category responses during the generalization phase. If participants had developed a response bias toward the compact category, we might have expected that they would make more compact category responses. Further arguments against the response bias account will be presented later.

A second alternative account of the results of Experiment 1 holds that categorization responses in that experiment were governed not by the similarity of the stimuli to the exemplars of the two categories but rather by the variability of the two categories. The compact distribution is wider than the dispersed distribution along the minor diagonal of the parameter space. The compact category is thus associated with a wider range of stimulus values in the vicinity of the test stimuli. Participants may have assigned the test stimuli to the compact category for this reason.