Running-head: Theories, ideals, and property induction Theories, Ideals, and Category-Based Property Induction

One important property of human object categories is that they are the categories to which newly observed properties are likely to be generalized. Another is that such categories often exhibit theoretical coherence. This study examined the relationship between inductive generalization and theoretical coherence by manipulating the causal knowledge associated with novel categories and assessing the resulting strength of property inductions. We found that the theoretical coherence afforded to category members by inter-feature causal relationships strengthened inductive generalizations. This effect obtained across a variety of causal network topologies, and across both natural kinds and artifacts. We propose an extension to Barsalou’s (1985) notion of ideals whereby an exemplar’s degree of category membership (and its propensity to support inductive generalizations) are increased not only when it satisfies human goals, but when it satisfies any of the complex causal and explanatory knowledge that people have about a kind. Theories, ideals, and property induction 3 Theories, Ideals, and Category-Based Property Induction Why do humans categorize the world the way they do and not otherwise? Intuitions inform us that there is something natural and sensible about categories like birds, apples, and diamonds, but something unnatural and strange about categories like weighs less than a ton, striped with more than one leg, and small blue triangle. Indeed, the latter examples do not seem to be categories at all. These intuitions are not based on mere experience with the categories, as unicorns and death stars seem like good categories even though they don’t exist, and we experience objects that weigh less than a ton everyday. Nor are they based on the presence of a single linguistic term denoting them: New World Squirrel Monkey is a separate species and Apple laptop computer is a common everyday object. Psychological research has enumerated some of the ways in which good or natural categories differ from arbitrary collections of objects. When asked to identify a new object, people will typically respond with a category label rather than an arbitrary class descriptor (Brown, 1958). For example, when a questioner points at a koala and asks “What’s that?,” people are likely to respond with “koala” but never with “weighs less than a ton.” Good categories also support confident inductive generalizations. Upon learning that a category member possesses a novel property, a category supports an inductive generalization to the extent that the observer concludes that many or all of the members of the category also possess the new property (Goodman, 1983; Kornblith, 1993; Shipley, 1993). For example, learning that one koala has a three-chambered heart would provide strong support for the conclusion that all koalas have three-chambered hearts, but no support at all for the conclusion that all things less than a ton have threechambered hearts. This is the case even though the koala is a member of the class of things that weigh less than a ton as much as it is a koala. Theories, ideals, and property induction 4 Properties of categories such as preferential classification and property induction are two important ways in which humans use conceptual systems in order to “go beyond the information given” to use their past experience to respond effectively to new situations. Classification allows individuals to reason from the general to the specific, by allowing unobserved features to be inferred on the basis of category membership (cf. Anderson, 1991). And, property induction supports reasoning from the specific to the general by allowing observed features to be projected to the category as a whole (and then to future category members). Not surprisingly then, establishing the origins of human categories that lead to those categories exhibiting preferential classification and strong inductive generalizations has been and remains a critical practical and theoretical issue. Perhaps the simplest answer to the question of the origins of human categories, and also the one that has dominated categorization research, is that people induce categories directly from what they observe. On this account, categories correspond to the information-rich clusters of features that occur naturally in the environment. This family resemblance view of categorization has been described as one where boundaries between categories are those that “cut nature at its joints” (Rosch & Mervis, 1975, also see Malt, 1995). The family resemblance view explains not only why koalas are classified as koalas and not “mammals” or “brown koalas,” (Corter & Gluck, 1992; Rosch, Mervis, Gray, Johnson, & Boyes-Braem, 1976) but also why they are not classified as “weighs less than a ton.” Objects that weigh less than a ton do not exhibit the family resemblance system of information-rich correlated features that “pick out” a natural category. More recently the family resemblance approach has been elaborated by a theorybased view of categorical structure. This approach emphasizes the importance of people’s conceptual background knowledge, in addition to statistical structure in the experienced environment, in a variety of category learning and categorization tasks. For Theories, ideals, and property induction 5 example, research participants asked to sort multi-dimensional stimuli exhibiting a family resemblance structure group those items into the family resemblance categories only when the categories’ features can be related to a common theme on the basis of participants’ prior knowledge. In the absence of such knowledge, objects are sorted on the basis of a single dimension. (Ahn & Medin, 1992; Kaplan & Murphy, 1999; Medin, Wattenmaker, & Hampson, 1987; Regehr & Brooks, 1995; Smith, 1981; Spalding & Murphy, 1996). The family resemblance view also predicts that categories should be easy to learn to the extent they are statistically simple, that is, linearly separable from one another. However, not only has research found that that non-linearly separable categories are as easy (and sometimes easier) to learn than linearly separable ones (e.g., Medin & Schwanenflugel, 1981), the rapid acquisition of new concepts that occurs during childhood does not seem to be reflected in the slow, labor-intensive acquisition of categories that characterizes learning of experimental categories (both linearly and non-linearly separable ones) employing meaningless stimuli. Instead, drastically accelerated category learning occurs only when the to-be-learned category is coherent in light of prior knowledge (Kaplan & Murphy, 2000; Murphy & Allopenna, 1994; Rehder & Ross, in press; Wattenmaker, Dewey, Murphy, & Medin, 1986). Finally, knowledge has also been found to have important impacts on classification behavior itself, both when prior knowledge in long-term memory is prompted (Rehder & Ross, in press; Wisniewski, 1995) and when it is taught as part of an experimental session (Ahn, 1998; Ahn, Kim, Lassaline, & Dennis, 2000; Rehder, 2001b; Rehder & Hastie, in press; Sloman, Love, & Ahn, 1998). As compared to category learning and categorization however, there have been fewer studies of the effects of background theories on category-based inferences such as the inductive generalization of new properties. The present research investigates the hypothesis that inductive generalization inferences are moderated or mediated by a category’s theoretical coherence. There are several reasons to expect that the knowledge Theories, ideals, and property induction 6 that underlies our representation of categories will influence inductive inferences. For example, according to one proposal, people assume that many categories (especially natural kinds) possess an underlying essence that makes an object the kind of thing it is. The essence is presumed to generate many of the properties of natural kinds that people observe (Atran, 1990; Coley, Medin, & Atran, 1997; Gelman, Coley, & Gottfried, 1994; Hirschfeld & Gelman, 1994; 1996; Mayr, 1988; Mayr, 1991; Medin & Ortony, 1989; Rehder & Hastie, in press). On this account, categories organized around an essence share a “deep resemblance”. The existence of such an essence explains many apparent correlations between the features associated with categories, and it leads to the belief that there are more yet-to-be-discovered features in common as well (Kornblith, 1993; Shipley, 1993). The presence of a causally-potent essence may promote stronger inductive generalizations because the essence is expected to generate many category features (Gelman & Coley, 1991; Gelman et al., 1994). A few prior empirical findings support the view that the background knowledge possessed by categorizers influences the inductive inferences they make. For example, Gelman (1988) found that second graders, but not preschoolers, were likely to generalize new properties to natural kinds than to artifacts. She attributes this result to socially-acquired folk theories about natural kinds that lead older children to believe they are homogenous with respect to not-yet-discovered properties (also see Lopez, Atran, Coley, Medin, and Smith, 1997, and Proffitt, Coley, and Medin, 2000, for adult studies of the influence of folk theories on induction). Carey (1985) attributed 4-yearolds’ stronger generalization of properties from people to other animals, rather than vice versa, to a conceptual organization in which people are the prototypical animal. Controlled research with adult subjects has also documented some effects of background causal knowledge about properties on the generalization of those properties from one exemplar to others. Smith, Shafir, and Osherson (1993; see also, Osherson, Smith, Myers, & Stob, 1994) found that when meaningful novel properties Theories, ideals, and property induction 7 were attributed to an exemplar, that simple similarity and typicality alone did not predict generalization of the property to other category members. For example, research participants judged that it was more likely that German Shepherds could bite through barbed wire when first told that poodles could bite through barbed wire as compared to being told that Dobermans could bite through barbed wire. This result held despite the fact that German Shepherds are more similar to Dobermans than to poodles. Intuitively, it seems that the participants were reasoning about the causal preconditions for the capacity to bite through barbed wire, and judged that a German Shepherd certainly could do so if a small dog like a poodle could but not if another powerful dog had the capacity. Sloman (1997; 1994) has also provided several demonstrations that the causes of a novel property moderate the generalizability of that property from one category exemplar to others. Although these studies are informative they are correlational in nature. For example, the co-development of theories about natural kinds and pattern of property inductions found by Gelman (1988), may be a result of the influence of theories on induction, or it may be due to a third underlying factor that affects both. The goal of the current study was to establish the causal influence of theories on property induction via experimental manipulations of the theories. This end was achieved by teaching undergraduates novel categories that either were or were not described as possessing causal knowledge that connected the categories’ features. Table 1 presents an example of the features and inter-feature causal relationships for one of the six novel categories used in the present research. Lake Victoria Shrimp were described to experimental participants as possessing four distinctive binary features and three inter-feature causal relationships among those features. In this example, the causal links formed a commoncause causal schema in which one feature (F1) was the common cause of the three remaining features (F2, F3, and F4). The knowledge associated with categories such as Lake Victoria Shrimp was intended to be a simplified analog of the essentialist Theories, ideals, and property induction 8 knowledge associated with real-world natural kinds. For example, the belief that shrimp DNA (for many people, the essence of animal kinds) causes shrimp features like having fins, locomoting by swimming, living in the sea, having a carapace, and so on. To determine whether categories with a common-cause schema support stronger inductive inferences, participants in the present experiments performed a series of property induction trials after learning about such categories. They were told about one category member that possessed a novel property, and were asked to estimate the proportion of other category members that possessed the novel property. Because such inductive inferences are also sensitive to the type of property, and the interaction between the type of property and category (Heit & Rubinstein, 1994; Nisbett, Krantz, Jepson, & Kunda, 1983; Thagard & Nisbett, 1982), we use blank properties that are unfamiliar to the participants (Osherson, Smith, Wilkie, Lopez, & Shafir, 1990). Table 1 includes the blank properties for Lake Victoria Shrimp. For example, participants are posed the hypothetical, "Suppose one Lake Victoria Shrimp has been found that has mucus that is slightly acidic," and then asked, "What proportion of all Lake Victoria Shrimp have mucus that is slightly acidic?" The predicate "mucus is slightly acidic" was deliberately chosen to be unfamiliar or "blank" to participants. Performance on the property induction task by participants who were provided with a common-cause causal schema is compared to a control group that was taught the novel category but without the associated knowledge (i.e., without the three inter-feature causal relationships). An additional goal of our design was to isolate the mechanism whereby the common-cause causal schema promotes inductive generalizations. For example, one possibility is that categories with a common-cause schema support stronger inductive generalization as a result of the global coherence afforded by inter-feature relations, regardless of the arrangement of the causal links. Alternatively, the specific form of the common-cause schema may promote inductions because it prompts categorizers to Theories, ideals, and property induction 9 assume that new features are also likely to be produced by the common cause. For example, when instructed on the Lake Victoria Shrimp category of Table 1 and a single category member that has acidic mucus, participants might reason that “high amounts of the ACh neurotransmitter causes many properties of Lake Victoria Shrimp, it might also cause acidic mucus as well.” To discriminate between these possibilities, the common-cause schema and the two additional causal schemas shown in Figure 1 were tested. Whereas in the common-cause schema one feature is the cause of the other three features, in the common-effect schema one feature (F4) is caused by the other three features (F1, F2, and F3); and in the chain schema, feature F1 causes F2, which in turn causes F3, which in turn causes F4. Only the common-cause schema will strengthen inductive generalizations if such inductions are based on the assumption that there is a single underlying generative cause, whereas all three schemas will promote inductions if they are a function a category’s global coherence, regardless of the specific arrangement of causal links. The three causal schemas in Figure 1 also enable a test of the hypothesis that categories are organized around a single central feature regardless of whether the feature is a cause or an effect (i.e., the common-cause and common-effect schemas) promote inductive generalizations, as compared to schemas without any such central feature (i.e., the chain schema). Indeed, the prior research cited earlier suggests that categories organized around a central theme have special properties regarding both category construction (e.g., Medin et al., 1987) and learning (e.g., Murphy & Allopenna, 1994). The importance of a central theme has also been suggested by prior work that has manipulated experimentally-provided causal knowledge. Rehder and Hastie (in press) found that features occupying a central position in a causal network dominated classification decisions. And, Ahn (1999) found that common-cause and common-effect schemas (but not chain schemas) elicit family-resemblance sorting. The current studies test the generality of such conclusions regarding a causally-central themes to the Theories, ideals, and property induction 10 inductive generalization of new properties. Although the main focus of the current research is on the effect of a category’s causal structure on property induction, some investigators have proposed that different cognitive mechanisms will be invoked depending on the kind of the category. For example, researchers have suggested that reasoning with biological kinds will be influenced by a domain-specific folk biology with which humans are innately endowed or which emerges early in development (Gelman & Kremer, 1991; Gopnik & Wellman, 1994). Keil (1995) has suggested that artifacts are also often understood in terms of the purpose they serve, a mode of understanding referred to as teleological-adaptive explanation (Bloom, 1996; 1998; Gould & Lewontin, 1978; Mayr, 1982; Rips, 1989). Indeed, Gelman's (1988) finding that elementary school children more often generalize new properties to natural kinds than to artifacts may have arisen from differences in the domain-specific mechanisms with which those kinds are processed, rather than differences in their background knowledge. To explore these questions, in the following experiments causal schema (common-cause, common-effect, or chain) was manipulated and crossed orthogonally with the kind of category. Specifically, we tested two biological kinds (Lake Victoria Shrimp, Kehoe Ants), two non-living natural kinds (Myastars and Meteoric Sodium Carbonate) and two artifacts (Romanian Rogos and Neptune Personal Computers), each with all four of the causal schema structures (common-cause, common-effect, chain, and no-cause control). Finally, the current study investigated the possibility that the effect of category coherence on property induction is mediated by the characteristics of the exemplar that displays the new to-be-generalized property. For example, a well-established research finding is that property inductions are strengthened when the new property is displayed by a typical versus atypical category member (Rips, 1975, also see Osherson et al., 1990). However, in the presence of causal knowledge about a category, exemplars that display new properties vary not only to the extent the which they are typical (in the Theories, ideals, and property induction 11 sense of having many shared features) of the category, but also how well they confirm or disconfirm their category’s causal laws. Indeed, prior research has found that goodness of category membership of an exemplar is enhanced for those exemplars that instantiate a category’s causal relationships (Rehder, 1999; 2001b; Rehder & Hastie, in press). This raises the possibility that causal schemas strengthen inductive generalizations, but only when the exemplar with the new property exemplifies the category’s causal theory. To assess this possibility, the present experiments also collected judgments of the goodness of category membership of those exemplars that displayed new to-be-generalized blank properties. As we will show, those category members that enhance the projection of new properties to a category are those that are theoretically ideal, that is, that manifest their category’s causal laws.