Elsevier

Acta Psychologica

Volume 164, February 2016, Pages 112-126
Acta Psychologica

Learning of role-governed and thematic categories

https://doi.org/10.1016/j.actpsy.2015.10.011Get rights and content

Highlights

  • ā€¢

    We examined how actors in events are categorized based on common relations.

  • ā€¢

    Relational categorization from highly similar events required little support.

  • ā€¢

    However when events only shared abstract relations, categorization was difficult.

  • ā€¢

    Analogical comparison enabled categorization based on common abstract relations.

Abstract

Natural categories are often based on intrinsic characteristics, such as shared features, but they can also be based on extrinsic relationships to items outside the categories. Examples of relational categories include items that share a thematic relation or items that share a common role. Five experiments used an artificial category learning paradigm to investigate whether people can learn role-governed and thematic categories without explicit instruction or linguistic support. Participants viewed film clips in which objects were engaged in similar actions and then were asked to group together objects that they believed were in the same category. Experiments 1 and 2 demonstrated that while people spontaneously grouped items using both role-governed and thematic relations, when forced to choose between the two, most preferred role-governed categories. In Experiment 3, category labels increased this preference. Experiment 4 found that people failed to group items based on more abstract role relations when the specific relations differed (e.g., objects that prevented different actions). However, Experiment 5 showed that people could identify them with the aid of comparison. We concluded that people can form role-governed categories even with minimal perceptual and linguistic cues.

Introduction

The psychology of concepts has primarily studied simple object categories, e.g., chairs and dogs, or their even simpler experimental stand-ins: colored shapes, dot patterns, or schematic animals (e.g., Posner and Keele, 1968, Rosch, 1973). Such categories are generally based on descriptive features intrinsic to category members, e.g., dogs have fur and four limbs, and that are organized into taxonomic classification hierarchies, e.g., dog, mammal, animal.

Recent research has given increasing attention to relational categories, whose membership is determined by extrinsic relationships among items (Gentner et al., 2011, Gentner and Kurtz, 2005, Goldwater and Markman, 2011, Goldwater et al., 2011, Kittur et al., 2006, Markman and Stilwell, 2001, Rein et al., 2010, Ross and Murphy, 1999, Tomlinson and Love, 2010). There are many kinds of relational categories, but they are similar in their dependence on context and extrinsic relations. For example, thematic categories contain objects that have a specific functional relationship, such as soup and bowl, or honey and bee. Here the items in the same category are not necessarily similar but have a complementary relationship: The soup is eaten out of the bowl; the bee makes the honey (Estes et al., 2011, Lewis et al., 2015, Lin and Murphy, 2001, Markman, 1989). Script-based categories group together objects that play the same role in a familiar event structure. For example, desserts are all eaten at the end of the meal, as a treat; greetings are the things we say at the beginning of an interaction (Nelson, 1986, Nguyen and Murphy, 2003, Ross and Murphy, 1999).

Similar to taxonomic categories, some of these relational categories also have overlapping intrinsic features. For example, desserts are usually but not always sweet. However, category membership is more strongly determined by the extrinsic relational properties. A cookie eaten before the main course is not considered dessert, and cheese or an unsweetened biscuit eaten at the end of the meal can be considered dessert. Someone who only attended to the intrinsic properties of cookies or pancakes would not be able to identify when they occurred as desserts.

The present article focuses on one type of relational category, role-governed categories (Goldwater et al., 2011, Markman and Stilwell, 2001). In such categories, membership is determined by the kind of role that the object plays relative to others. For example, a guest is a person who stays in someone else's home; a teacher is someone who instructs someone else; and a barrier is something that prevents access to a goal. Although the members of these categories can be similar, the critical determinant of category membership is the exemplar's role in some larger event or structure. A wealthy, domineering person who orders everyone around is not a CEO unless she actually has the proper role relative to others within a corporate structure. Many homes are houses, but a house no one lives in is not a home. Goldwater et al. (2011) empirically demonstrated the link between role-governed categories and relational structures and differentiated them from taxonomic feature-based categories along a number of dimensions (see also Gentner and Kurtz, 2005, Rein et al., 2010). However, the initial formation of role-governed categories has been studied to a much lesser degree (see Gentner et al., 2011, Goldwater and Markman, 2011, Jones and Love, 2007, for how common roles affect categorization and similarity of familiar items). The initial formation of role-governed categories is the focus of the current work.

It will be important for our purposes to distinguish such role-governed categories from thematic categories. Thematic categories include two or more kinds of objects that are connected by a specific relation, such as leash-and-dog. Note that these two objects do not have a common role, but instead play complementary roles to one another: The leash holds the dog; the dog is held by the leash. In contrast, objects in role-governed categories have the same roles in relationship to other objects not in the category. For example, all barriers impede access to a goal; but goals are not in the barrier category.

This example points out that there is often a close relationship between thematic and role-governed categories: One half of a thematic category often forms a role-governed category with similar items. For example, a teacher and student might form a thematic pair. Teachers themselves form a role-governed category, because they all share the same relationship to their students. Students form a role-governed category, because they all share the same relationship to their teachers. Similarly, a barrier and goal are linked thematically, and they are also each a member of different role-governed categories. This linkage between the two types of categories does not always hold. Although leashes and dogs might form a thematic category, dogs cannot be said to form a role-governed category, because they are not primarily defined in terms of their relationship to other things.

At this point, some readers may be wondering whether all these groupings really are categories. Not only do they group things that are not at all similar, like leash and dog, they include a diverse set of relationships as the bases for classificationā€”thematic relations, scripts, roles, and similarity. As discussed by Murphy (2010), identification of these kinds of categories is not just a theoretical claim but an empirical observation of the kinds of classifications that people actually make. People group items together in these ways when asked to form categories or to group together the same kinds of things (Estes et al., 2011, Gentner and Brem, 1999, Goldwater and Markman, 2011, Lin and Murphy, 2001, Ross and Murphy, 1999). They can influence memory performance (Rabinowitz & Mandler, 1983). Furthermore, relational categories provide a basis for category-based inferences (Rein et al., 2010, Ross and Murphy, 1999). Thus, such groupings share many of the functions of familiar taxonomic categories.

Another important connection to role-based categories is the creation or comprehension of analogies. Analogical inference is based on shared relational structure (Gentner, 1983, Holyoak and Thagard, 1989). The classic example of comparing the structure of the atom to the solar system works not because the sun and its planets are similar to the parts of the atom, but because the pattern of relationships between the sun and its planets is similar to that of the relations between an atom's nucleus and its electrons. However, analogies are not trivial to notice on one's own, and much research has focused on why analogical reasoning can be so difficult (e.g., Gick and Holyoak, 1980, Keane, 1997, Rottman et al., 2012). This difficulty suggests in turn that role-based categories may be difficult to identify as well. We explore this possibility in the next section.

Much early work on role-governed categories and relational categories more generally has focused on distinguishing familiar role-governed categories from familiar taxonomic categories. Exemplars of role categories such as barrier are less similar to each other, e.g., wall and poverty, than are exemplars of taxonomic categories such as fruit (Gentner & Kurtz, 2005). Role-governed categories such as friends elicited extrinsic properties about how category members relate to others, like ā€œthey are always there for you,ā€ while taxonomic categories such as knife, elicited properties intrinsic to category members, like ā€œmade of metalā€ (Goldwater et al., 2011). In addition, relational categories' meanings are more contextually mutable (Gentner & Asmuth, 2008).

These properties of role-governed categories seem likely to make them harder to learn than more familiar taxonomic categories. Just as people find it difficult to use the solution from one problem to solve another problem that is relationally identical but superficially different (Gick & Holyoak, 1980), they may find it difficult to identify a common role like guest. If one did not already know this word, would one identify people as being guests across all the different settings in which it is used? Would they identify houseguests, hotel guests (who pay for their services), people invited to a restaurant for dinner (who eat for free), guests at a wedding (who are not served food), or perhaps even a dragonfly visiting a garden (an attested example from Goldwater et al., 2011) all as being the same kind of entity, abstracting across their ages, genders, social classes, and even species? Could they identify the common role across very different events? In contrast to thematic categories, which often involve associated objects, identifying role-governed categories can require considerable abstraction. The fact that we already know the word guest makes the similarity of these different entities easy to detect, but we will shortly review literature that suggests that detection of novel role-governed categories is not so easy. That is, perhaps exposure to the word guest is important for our learning that category.

The category of guest co-occurs with another role-governed category, host, and the two form a thematic category. Much research now shows that many people will classify thematic pairs such as guests and hosts as being in the same category or as ā€œthe same kind of thingā€ (Estes et al., 2011, Goldwater and Markman, 2011, Lin and Murphy, 2001, Murphy, 2001). In light of that, it is not clear that people will reliably detect role-governed categories themselves in the absence of a linguistic label or an explicit search for commonalities. That is, they might think that a host and his guest go together but, prior to learning the word host, may not perceive that two different hosts are the same kind of person. The host bears a direct relationship to his or her guest but not to a different host.

Indeed, relational categories are not always correctly identified as such. For example, consider Keil and Batterman's finding (see Keil, 1989, ch. 4) that children misidentify the relational term uncle to be a feature-based category (a nice man who comes to visit and gives you presents) until they are in 4th grade. Clearly, these common features are easier to acquire than the familial relation that is the true basis of the word. Gentner and Kurtz (2005) point out that relational words are generally learned later, that it takes longer to produce examples of relational categories, and that people who view such examples (barriers, filters, guides, etc.) often have difficulty identifying what they have in common. Such results raise the question of whether people can identify and use relational categories as they do taxonomic categories, without instruction. Indeed, after considering this question, Gentner and Kurtz (2005 p. 165) propose that learning relational categories is at least in part ā€œthrough cultural guidance: in particular, through linguistic labels.ā€ Gentner et al. (2011) showed that 4ā€“5-year-old children identify relational categories based on a single example much better if they are given linguistic support, for example, ā€œthis knife is the dax for the watermelon.ā€ Such sentences not only label the category but use relational language that makes the nature of the category clear.

Another aid to role-governed category learning is comparison of category exemplars. Goldwater and Markman (2011), using linguistic stimuli, showed that without the support of a category label or explicit comparison, people were four times more likely to group thematic pairs, such as bodyguard and celebrity, than they were to group things that played the same role across domains, such as bodyguard and force-field (both are protectors) (see also Gentner et al., 2011, Jung and Hummel, 2011). These results are similar to, and perhaps have the same psychological basis, as the finding that analogies are difficult to notice without explicit instruction (Gick and Holyoak, 1983, Goldwater and Gentner, 2015, Catrambone and Holyoak, 1989).

In sum, it is not clear just how well people can identify and learn these relational categories under neutral conditions without support from language, a teacher, or a helpful setting. We present a series of studies examining under what conditions people spontaneously form novel role-governed categories from viewing novel events. We contrasted such categories with novel thematic categories, which are more directly associated. To construct novel categories of both types, we created events in which different kinds of objects interacted. Participants viewed films of objects that engaged in similar actions across the films. Some of the objects could be grouped into thematic pairs such as one chasing the other. However, across films, objects could be grouped into role-governed categories such as the chasees. We tested each kind of category independently (following Nguyen & Murphy, 2003) to identify whether people could identify role-governed categories, thematic categories, or both, from the same events. We also included conflict trials, in which one type of category had to be selected over the other, to identify which kind of category is more salient.

As mentioned earlier, role-governed categories often share attributes, which makes sense given that the ability to carry out a role may require certain properties. For example, teachers are usually adults with a college education; barriers are often bulky. Across experiments, we varied the similarity of the members of the role-governed categories. Because the paradigm required people to remember a number of novel objects and what they did in the films, there were always some perceptual cues to allow participants to remember which kinds of objects did what (we discuss the role of perceptual features in relational categories in the discussion). Since thematic categories involve grouping different kinds of objects, these categories did not generally have as much perceptual similarity, just as leash and dog or soup and bowl are not similar. However, these pairs had consistent relationships across the films. For example, a green object might always chase a red object, and participants who attended to the thematic relations could then group green and red objects into thematic pairs at test.

Section snippets

Experiment 1

In order to discover whether people identify and use novel role-governed categories to classify objects, we constructed film clips in which objects interacted. In each clip, there were two objects that interacted in meaningful ways. In one set of three clips one object crushed the other one; in another set, one object chased another one; in the third, one object lifted another. (These descriptions are shorthands for the specific events that occurred in the clips. The actual actions had

Experiment 2

Participants clearly identified the roles and thematic relations in Experiment 1 and used them to classify the objects. There was a general preference for role responses, but it is not clear whether this is due to a pure preference for roles over thematic relations, as the role-governed categories shared common shapes as well as common roles. This is probably naturalistic, as things that do the same actions and engage in the same relations must have common parts and other properties in order to

Experiment 3

Ellen Markman (1989) has discussed the paradoxical finding that children often make thematic groupings but that their category names are generally not thematically based. She suggested that children have a taxonomic assumption to interpret nouns as being feature-based. This makes sense, as objects that have complementary relations generally have different properties, reflecting their different roles. Soups and bowls differ greatly in their origins, functions, parts, and constituent ingredients.

Experiment 4

The same-role and thematic pairs tested in the previous experiments were all based on very similar actions within a set of clips. The chasing or lifting or crushing were all quite similar, even though they differed in their location on the screen, the shape and specific motions of objects carrying them out, and so on. Each event category can be represented by first-order relations that directly relate objects, e.g., one object closely follows another. Some roles we identify in everyday life are

Experiment 5

People's difficulty in identifying common abstract roles in Experiment 4 seems related to the general finding that people find analogy difficult when superficial features and first-order relations differ across items (e.g., Gick & Holyoak, 1980). In one preventing clip, a character stops another character from reaching a balloon; in another, the preventer stops another character from exiting through a gap in the wall. Although both are abstractly good examples of preventing, they involve

General discussion

There is growing evidence for the prominence of role-governed categories in our conceptual system. This article is first to show that people can abstract novel role-governed categories from viewing novel events, without tuition or guidance from linguistic labels. Experiment 1 and Experiment 2 found that people can readily detect role-governed categories as well as thematic categories based on first-order relational commonalities. Experiment 3 showed that category labels shifted focus from

Acknowledgments

The research was supported in part by NSF grant BCS 1128769. ARC grant DP150104267 supports the research of the first author. We are grateful to Arthur Markman and Mitchell Rabinowitz for helpful comments.

References (64)

  • D.G. Kemler Nelson

    Principle-based inferences in young children's categorization: Revisiting the impact of function on the naming of artifacts

    Cognitive Development

    (1995)
  • B. Landau et al.

    The importance of shape in early lexical learning

    Cognitive Development

    (1988)
  • A.B. Markman et al.

    Structural alignment during similarity comparisons

    Cognitive Psychology

    (1993)
  • E.M. Markman et al.

    Children's sensitivity to constraints on word meaning: Taxonomic vs thematic relations

    Cognitive Psychology

    (1984)
  • D.L. Medin et al.

    Folkbiology of freshwater fish

    Cognition

    (2006)
  • E.H. Rosch

    On the internal structure of perceptual and semantic categories

  • B.H. Ross et al.

    Food for thought: Cross-classification and category organization in a complex real-world domain

    Cognitive Psychology

    (1999)
  • S.S. Smiley et al.

    Conceptual preferences for thematic or taxonomic relations: A nonmonotonic age trend from preschool to old age

    Journal of Experimental Child Psychology

    (1979)
  • D. Billman et al.

    Unsupervised concept learning and value systematicity: A complex whole aids learning the parts

    Journal of Experimental Psychology: Learning, Memory, and Cognition

    (1996)
  • R. Catrambone et al.

    Overcoming contextual limitations on problem-solving transfer

    Journal of Experimental Psychology: Learning, Memory, and Cognition

    (1989)
  • S. Day et al.

    Nonintentional analogical inference in text comprehension

    Memory & Cognition

    (2007)
  • S. Day et al.

    Analogical transfer from a simulated physical system

    Journal of Experimental Psychology: Learning, Memory, and Cognition

    (2011)
  • L.A. Doumas et al.

    A theory of the discovery and predication of relational concepts

    Psychological Review

    (2008)
  • D. Gentner

    Why we're so smart

  • D. Gentner

    Bootstrapping the mind: Analogical processes and symbol systems

    Cognitive Science

    (2010)
  • D. Gentner et al.

    Can relationality be distinguished from abstractness in noun mutability?

  • D. Gentner et al.

    Is snow really like shovel? Distinguishing similarity from thematic relatedness

  • D. Gentner et al.

    Relational categories

  • D. Gentner et al.

    Structure mapping in analogy and similarity

    American Psychologist

    (1997)
  • D. Gentner et al.

    Structure mapping and relational language support children's learning of relational categories

    Child Development

    (2011)
  • R.L. Goldstone et al.

    Time course of comparison.

    Journal of Experimental Psychology: Learning, Memory, and Cognition

    (1994)
  • M.B. Goldwater et al.

    Categorizing entities by common role

    Psychonomic Bulletin & Review

    (2011)
  • Cited by (15)

    • Evidence for multiple sources of inductive potential: Occupations and their relations to social institutions

      2021, Cognitive Psychology
      Citation Excerpt :

      The specificity and richness of institutional roles also opens up new avenues for future research. The literature on role-governed categories has focused on category formation (Goldwater, Bainbridge, & Murphy, 2016); categories like guest and leader require participants to identify relational similarities across diverse contexts and suppress alternative feature-based representations. Therefore, the hard part is engaging in the abstraction or analogical reasoning needed to isolate the roles; there is less need for protracted learning about the relational structures per se.

    • Preschoolers and multi-digit numbers: A path to mathematics through the symbols themselves

      2019, Cognition
      Citation Excerpt :

      Number names and written digits comprise two parallel relational structures of the kind studied within the framework of Structure Mapping (Gentner, 1983, 2010; Gentner & Colhoun, 2010. Given varying but relationally alignable surface formsā€”for example, models of the solar system and atoms or relational series such as big-little-bigā€”learners can discover the relational structure and apply that structure to new instances (Gick & Holyoak, 1983; Goldwater, Bainbridge, & Murphy, 2016; Loewenstein, Thompson, & Gentner, 1999). Moreover, research with children as well as adults indicates that relational structures may be learned from exposure to two alignable series, without explicit teaching or feedback (Christie & Gentner, 2010; Fisher, 1996; Gentner et al., 2016; Namy & Gentner, 2002).

    • Influence of category coherence and type of base-rate acquisition on property generalization

      2017, Acta Psychologica
      Citation Excerpt :

      Future studies should compare diverse forms of base-rate information to understand category coherence and inferences, similar to Gottlieb et al.'s (2007) and Hertwig et al.'s (2004) comparative studies. Moreover, by examining the interaction between coherence for ad hoc categories (e.g., things to sell at a garage sale; see Barsalou, 1983; Goldwater, Bainbridge, & Murphy, 2016; Ratneshwar, Barsalou, Pechmann, & Moore, 2001) and acquisition of base-rate information, the results of this research can be expanded. Further research is necessary to determine the effect of category members' gender on the perception of category coherence.

    View all citing articles on Scopus
    View full text