Capacities versus Performance Models

Capacities versus Performance Models
One aspect of this argument (although not the only one) involves
discriminating two diVerent components of behavior. One we can call
capacities, namely people’s general abilities. For example, when we study
memory capacity or language competence, we may be attempting to
characterize people’s knowledge or representational structure in a general
way. The second component, which I have given the awkward name
performance models, refers to what people actually do in specific
circumstances. Performance models depend not only on the underlying
abilities, but also on the situations that people find themselves in. Using a
linguistic example, it is possible that I could figure out the meaning of a
center-embedded sentence such as The cat the dog the mouse chased feared
ate, but I would do so by trying to divide the sentence into clauses and
match the subject with its verb, hopefully with pen and paper. However, in
real life, no one ever says such sentences to me, and I never do process them.
(If someone did slip one into a real conversation, the chance of my
interpreting it correctly in real time would be slim.) Thus, a theory of my
language understanding that does not provide for comprehension of doubly
embedded sentences might correctly account for my use of language outside
of experimental tasks—it would be a performance model of my actual
language use.
2
The study of language capacity would, on this definition, include basically
anything that anyone could do with linguistic materials, from writing poetry
to solving anagrams to cross-modal priming with a fast response deadline.
2
This distinction sounds ominously like the competence/performance distinction in
generative linguistics. However, that somewhat dubious distinction is usually intended to focus
on underlying knowledge (competence) versus less interesting processes that filter the
knowledge in external behavior. In contrast, I think that psychological studies of both
capacities and performance models fall on the performance side of this distinction, because both
are about mental representations and processes, and not just underlying knowledge. Also,
whereas the competence/performance distinction is usually drawn to allow linguists to ignore
performance data, I am suggesting that performance models are the more interesting topic of
study. In short, despite the superficial similarity, try not to think of the competence/
performance distinction when reading this section.
14 Gregory L. MurphyIn contrast, a performance model of language would attempt to answer the
question of how people deal with utterances of the sort they normally hear
and how they produce the utterances they normally utter, in the contexts
that they normally do these things. Clearly, the two are very related in most
cases, but they also can deviate in others (e.g., the anagram case ... and
perhaps the cross-modal priming with the fast deadline?). In developing an
explanation of how people actually comprehend ambiguous words, say, one
is also a fortiori making claims about language capacity. If you argue that
people consider all the diVerent possible meanings of a word, you are
making a claim about people’s abilities. However, the reverse is not
necessarily the case. One may find evidence of cognitive capacities that do
not in fact participate in the normal behavior of that domain. The reason
could be that the experimental task required use of a capacity that is
normally not used due to its diYculty. If the anagram is really hard, you
may have to engage in diYcult strategies, such as consciously generating
words, writing down letter combinations, and thinking about spelling rules.
Another reason is that everyday life does not present a situation that would
benefit from such a capacity. I understand normal English sentences well
enough that I don’t need pen and paper to figure them out. Thus, my ability
to understand center-embedded sentences with pen and paper is simply not
relevant to my normal language use.
There are arguments to be made for why we should try to understand
cognitive capacities. I am not going to make those arguments. The next
section argues at length for why we need a performance model of concepts.
Here I will simply point out that if we want a theory of how people perceive
and behave in their everyday lives, we must not be too hamstrung by data
about their capacities. Clearly, any limitation on cognitive capacities will
apply very broadly; if short-term memory has room for only three or four
items even under favorable circumstances, we should not claim that people
have a dozen items in short-term memory during our category-learning task.
However, the fact that people can do something in certain laboratory
settings does not mean that they do do it in a particular, perhaps less
demanding situation. To find out whether they do, we need evidence from
within that setting and not from a very diVerent one.
The implications of this argument to the psychology of concepts are clear.
Because of the problem of unconstrained category construction, we can
make up categories that are very diYcult for subjects to learn or that are
peculiar. Clearly, the results of such studies tell us about cognitive
capacities—people’s ability to learn, remember, and make decisions. What
is not clear is whether these results are relevant to a performance model of
actual category learning. The fact that people may memorize exemplars to
learn some categories does not imply that they do so in real life unless the
Ecological Validity and the Study of Concepts 15real categories are like the experimental categories in certain respects. The
fact that people can learn categories consisting of orthogonal items (as in
Shepard et al.’s type VI task), and that this is harder than learning categories
that are almost orthogonal (e.g., type V), is also a fact about human
cognition. But it may or may not be a fact about human concept learning.
Whether it depends on the nature of actual concepts, and that is something
that an experiment using artificial materials cannot tell us.
Skeptics may question whether we should be trying to focus on
performance models rather than capacities. Isn’t it just as important to
understand the basic processes of the mind as it is to understand whatever
processes are used in most real behaviors? I address this question next.