Rosch, E. & Mervis, C.B. (1975).  Family resemblances: Studies in the internal structure

of categories.  Cognitive Psychology, 7, 573-605.

Purpose:     

A central theme of this article is that category membership, contrary to what many people purportedly believe, is not all-or-none.  If all categories had  an all-or-none structure, then why do people so easily recognize some members of a category as better, or more prototypical, than other members?  The authors report that in and of itself, recognizing that many categories have graded membership does not provide a satisfactory explanation for how categories arise.  The primary purpose of this article was to explore what the authors believed to be a major structural principle that governs the formation of categories, namely, family resemblance.  A category has a family resemblance structure when each category member shares several elements (or features, or attributes) with other members of the category, but there is not necessarily any single element that is common to every single member of the category.  Thus, categories are viewed  as networks of overlapping elements, none of which must be common to all examples of the given category.  The main questions under investigation were whether prototypical items are rated as such because they share many elements with other members of their own category (i.e. bear a greater family resemblance) and also because they have few elements in common with members of other categories (i.e. maximally discriminable from other categories)

Three different kinds of categories were investigated: superordinate level categories (e.g. furniture), basic level categories (e.g. chair) and artificial categories.  Superordinate and basic level semantic categories appear in everyday speech, albeit at different levels of abstraction.  Superordinate categories have as their members basic level category names, and basic level categories in turn contain object names.  For each type of category, the authors tested two aspects of the family resemblance hypothesis: 1) that the most prototypical members of a category are those that share the most attributes in common with other members of that category, and 2) are those with the least attributes in common with other categories.  Inclusion of experiments examining the structure of both kinds of categories serves to show that the proposed relationship between prototypicality and family resemblance is present in categories that people use everyday.  However, the authors recognize that in order to demonstrate that greater family resemblance leads to greater perceived prototypicality, it is necessary to use artificial categories with which people are not already familiar.

Experimental Work:

Superordinate Categories:  Experiments 1 and 2 sought to demonstrate that the "categorical" nature of superordinate level categories is due primarily to family resemblance, and not to any criterial attributes common to all members. 

In Experiment 1, the authors demonstrated that members of a superordinate category (which consists of basic level category names) that share many attributes with other members of the category are also rated as the most prototypical.  The number of shared attributes was obtained by simply presenting subjects with basic level category names and having them list as many attributes for it as they could.  Attributes were weighted according to the number of members in the category to which the attribute applied.  The sum of the attribute weights for a given item was considered a measure of family resemblance.  Prototypicality was essentially measured by asking people how prototypical a given item was of a category.   Simply put, members of a superordinate category that share many attributes with other members of the category are also rated as prototypical.   It should also be noted that two out of the six superordinate categories tested did not have a single attribute that applied to every member, and the remaining four categories had only one such attribute each.

Experiment 2 demonstrated that prototypical members of a given superordinate category are not prototypical of other categories.  A single measure of the degree to which a given item fit into one category and not others was obtained by presenting participants with basic level names and having them list three categories to which the basic level names could belong.  Categories listed first, second or third were given weights of 3,2 and 1 respectively.  The category dominance score was computed by subtracting weighted scores of the first and second most frequently mentioned superordinate categories from the designated superordinate.  This measure correlated strongly with prototypicality ratings.  Thus, prototypical members of a superordinate category are those that are unlikely to be representative of other categories.  The findings of experiments one and two are consistent with the two part family resemblance hypothesis that the authors proposed.

Basic Level Categories:    Experiments 3 & 4 were analogous to the first two, except that basic level categories were under investigation.  Experiment 3 demonstrated that members of a basic level category that shared many attributes with other members of the category were rated as prototypical.  Prototypicality ratings and family resemblance measures were obtained in almost exactly the same way as in Experiment 1, except that participants were shown pictures of objects, and rated the prototypicality or listed attributes of the specific object shown.  Experiment 4 demonstrated for basic level categories what Experiment 2 did for superordinate.   In both  cases, prototypical items at a given level of abstraction are maximally discriminable from items in other categories at the same level of abstraction.  For a given basic level category however, it was possible to obtain contrast categories (i.e. categories that share many physical and/or functional attributes with the target category, e.g. chair and sofa).  To obtain contrast categories, participants were to pretend that they were playing a guessing game, in which they were attempting to guess a word based on a list of attributes (the participants were not actually presented with a list of attributes, they had to pretend they had been).  The participants were told to imagine that their first guess, which they were given, was incorrect (e.g. chair). Their task was then to provide a second guess (e.g. sofa).  Attribute lists were then obtained for the contrast categories.  A negative correlation between prototypicality and number of shared attributes with items from the contrast category was obtained.  Thus, Experiments 3 & 4 confirmed that at the basic level of abstraction, highly prototypical items share many attributes with members from within the given category, and share few attributes with members from another category in direct contrast.

Artificial Categories:  The first four experiments were inherently correlational in nature, and thus do not provide evidence that high family resemblance and high discriminability are what causes items to become category prototypes.  Hence for Experiments 5 & 6, artificial categories were constructed in which the degree of family resemblance within a category and overlap of attributes across categories were manipulated.  In both experiments, participants learned to discriminate six-item categories consisting of unpronounceable strings of letters and numbers.  Subsequent to learning, reaction time to make category judgments and prototypicality ratings were measured.  In Experiment 5, the contrast sets did not overlap and only family resemblance was varied.  Results showed that, compared to items low in family resemblance, items with high family resemblance were learned faster, categorized faster after learning and rated as more prototypical.  Importantly, categories in which all items had equal family resemblance yielded no differences in terms of learning rates, categorization or prototypicality.

In Experiment 6, the degree to which contrast sets overlapped was varied, as was the degree of family resemblance.  For the control set, each item in the target category bore equal family resemblance to other items from that category, but varied in terms of how many attributes overlapped with members of the contrast category.  In Experiment 5, when contrast sets did not overlap and family resemblance was held constant for all items in a given category, rate of learning, response time to categorize, and prototypicality were equivalent for all items.  In contrast however, in Experiment 6, items that overlapped less with the contrast category were learned faster, categorized faster subsequent to learning, and rated as more prototypical.  The other contrast set varied both in terms of family resemblance and overlap, such that they opposed each other for some items, but not for others (i.e. one item was high in family resemblance and low in overlap, another was high in family resemblance but also high in overlap etc.).  Again, learning rates, response times, and prototypicality ratings of items were affected by the degree of overlap with the contrast set, even for items high in family resemblance, such that when family resemblance and overlap were opposed, they tended to cancel each other out.

Conclusions: 

The results of these experiments demonstrate that items viewed as prototypical of a category are those that share many common attributes with other members of the category and are highly discriminable from items in other categories.  The authors contend that these results are important in six ways in that they suggest: 1) the importance of family resemblance structure as a basis for category formation, 2) how prototypes and cue validity are complementary, 3) a structural rationale for the use of proximity scaling in the absence of definable category dimensions, 4) how prototype formation can be viewed as part of a general process by which categories are formed, 5) a new link between adult and children’s modes of categorization and 6) there is a concrete alternative to criterial attributes in understanding the logic of category structure.

Points for Discussion:

One of the nicest features of this series of studies is the simplicity.  But, one might argue that the simplicity makes it hard to generalize the results.  For example, at most, only six natural categories were examined in any experiment, and in experiment  4, only two target categories were used.  Furthermore, each category contained as members things that are objects.  So, what reason is there to believe that these results will generalize to concepts such as “fair game”, or “justice”?

Rosch argues throughout the article that category formation is not arbitrary, and that the basic level category represents the cognitive systems “carving of nature at its joints”.  That is, in nature there is a correlational structure that serves as the basis for prototype and in turn category formation.  However, the authors never specify exactly what features we perceive as correlated, nor do they offer any rationale for why some features are important for some categories but not for others.  How might these cause problems for the authors’ interpretation?  Recall that in the artificial categories used in experiments 5 & 6 the family resemblance structure was provided in the input and reinforced throughout learning.  But, does this necessarily mean that the family resemblance structure is also always available in naturally occurring categories?  It seems circular to answer “yes” to this question based on the fact that participants report attributes that have a correlational structure. 

This page was last updated:
07/18/2006 00:36