The Underlying Factors That Shape Our Concept Categories
Prototype & Other Effects on Classification
It is now widely recognized that the majority of conceptual category assessments are not all-or-nothing affairs. Some members of a category are judged better members of the category than others (note 1). Robins and sparrows are prototypical examples of birds, while penguins and chickens are not. Other examples, like hawks, seem to fall on a graded scale between the two extremes. Yet all are unequivocally birds. Bats, on the other hand, seem a lot like birds, but are not. Most people even judge small numbers like 2, 4 and 6 to be better examples of an even number than numbers like 4378 or 916.
Although prototype effects have proven difficult to quantify consistently, and are variously modeled in the literature, the effects are readily demonstrable. Consider just two characteristics of birds, namely size and flying ability. We can simply assign potential attribute values for bird size ranging from small, medium, large; and for flight as can fly and can’t fly (note 2). Then prototypical birds correspond to those that are small and can fly (e.g., finch, robin, jay). If we order the attribute values such that they increasingly depart from the prototype values, the corresponding ordering of exemplars agrees fairly well with experiments in which subjects grade them as increasingly poorer examples of birds:
Moreover, it is easy to see how a small flying mammal, like a bat, could be mistaken for a bird. Additional information, such as the salience (eg, prominence, frequency, and familiarity of observed attribute values) and the diagnosticity measures included in most concept models, will further bias the proto-typicality ratings, but the simple presence/absence of just a few characteristics is enough to explain much of the graded effect.
Multiple Category Functions
The source of the tension between consistent conceptual categories and prototype effects may stem from the interplay between three of the primary functions served by conception — namely, recognition, prediction and classification.
In recognition, we are attempting to match an observation with a learned category, for the purpose of shaping a timely and appropriate response. Noticing certain salient characteristics, like stripes or spots, triggers different concepts, like tiger or cow. Recognizing a tiger then produces a different response from recognizing a cow. This facility for quick and suggestive recognition based on appearances is of great survival value, and our conceptual representations seem tightly integrated with perception to accomplish this function.
In prediction, we use learned knowledge about concepts to anticipate unobserved or future characteristics. Once we have concluded something is a cow, we expect it will produce milk even if we have never observed this cow before. Such expectations often guide us in fulfilling our goals. In the most difficult cases, we need to know what something is, rather than merely how it appears, in order to make reliable predictions based on causal principles.
Classification serves yet a different purpose. Classifying concepts helps us organize our knowledge into a useful framework supporting deliberative reasoning, as opposed to simple association used for perceptual recognition and prediction. If we are familiar only with tigers, but have been told that both tigers and lions are big cats, we tend to infer that lions share some significant characteristics with tigers. Classification depends critically on what something is, as opposed to how something appears, and the characteristics relevant to classification are usually latent, rather than salient. In fact, conceptual recognition and prediction seem almost like elaborations of integrated perceptual detection/recognition/prediction processes, and may have arisen as a first evolutionary step toward the more sophisticated functions of classification, which are often far removed from perception.
Multiple Category Conditions
These kinds of overlapping conceptual functions may help explain why we often think ambivalently about categories — why we know ostriches are birds but rather poor examples of birds. Certain characteristics (e.g., bird DNA, as an identity condition, or feathered, as a normalcy condition) are key to category membership, but various other characteristics (e.g., nesting in trees, as a stereotypical condition, or ability to fly, as a cue condition), are only suggestive. We seem to maintain an operational separation between these different kinds of characteristics. If so, prototypical intuitions may arise from weighing and combining evidence from the more peripheral characteristics, for and against, to arrive at an overall degree of goodness of fit for a category (or an individual conceived of as a category of one).
The total number of relevant peripheral characteristics that match and don’t match certainly contributes, but high salience also correlates with prototypical assessments. For example, consider the characteristic of an apple’s color. To a first approximation, an apple’s possible color values may register as red, green, yellow, and brown (for rotten apples). Such color values reflect experiences with apples, and accordingly we have likely acquired some knowledge of the natural frequency distribution of these colors. Red apples are the most common, making them prototypical, followed by green and yellow, which score lower. Brown apples are like ostriches — members of the category, but decidedly atypical. When we see an untypical example, we can usually still classify it as an apple, but its prototypicality will be downgraded by its atypical color value.
Alternately, category membership questions more reliably depend on the satisfaction of core identity and normalcy characteristics, where identity characteristics require unequivocal satisfaction and any normalcy characteristics that are not satisfied must be justifiably explained away. Thus a chicken, satisfying the core identity and normalcy characteristics for a bird, is decidedly a bird, but it scores low on the weighting of peripheral characteristics for birds, creating a certain tension for category membership.
Two other possible sources of prototype effects relevant to most concept models concern the applicability and the diagnosticity of a characteristic. Oftentimes it can be difficult to decide if a characteristic applies to an instance of a concept. Having feathers is normal for birds, but if an exemplar has a body covering that cannot reliably be called either feathers or nonfeathers, its membership in the bird category will be biased according to the degree of resemblance to feathers. It is questionable then whether the characteristic applies. Although such situations might loosely be included in prototype effects, they are probably better classed as issues of confidence. Likewise, characteristic values that are highly diagnostic for a category provide a stronger bias toward membership than somewhat less diagnostic values.
Finally, relative diagnosticity may have a wider scope than basic prototypicality, since proper diagnosis may involve differential comparisons among all alternatives in a concept’s contrast set — e.g., for bird, the contrasting vertebrate concepts of fish, mammal, reptile, and amphibian. In other words, the characteristics of fish, mammals, reptiles, and amphibians may require consideration in ultimately deciding if something is a bird in borderline cases.
All in all, appreciating the basis of our concept categories can be an insightful, if messy, business. But it tends to undercut confidence in our intuitively clear concepts about many things.
Footnotes:
[1] Rosch, E. (1978). Principles of Categorization. In E. Rosch, & B. B. Lloyd, Cognition and Categorization (pp. 27–48). Hillsdale: Lawrence Erlbaum Associates.
[2] A more realistic account would also distinguish very small (eg, hummingbird) & very large (eg, ostrich), and perhaps awkwardness (eg, albatross), but this is not necessary to demonstrate prototype effects.
