storms-deboeck-ruts-2001.doc - Categorization in natural concepts 1 CATEGORIZATION OF NOVEL STIMULI IN WELL-KNOWN NATURAL CONCEPTS A CASE STUDY Gert

storms-deboeck-ruts-2001.doc - Categorization in natural...

This preview shows page 1 out of 28 pages.

You've reached the end of your free preview.

Want to read all 28 pages?

Unformatted text preview: Categorization in natural concepts 1 CATEGORIZATION OF NOVEL STIMULI IN WELL-KNOWN NATURAL CONCEPTS: A CASE STUDY Gert Storms Paul De Boeck Wim Ruts University of Leuven Mailing address: Psychology Department, University of Leuven, Tiensestraat 102, B-3000 Leuven, Belgium. Telephone number: (int) 32-16-326.088. E-mail address: [email protected]). Suggested running head: Categorization in natural concepts Categorization in natural concepts 2 Abstract This study investigated to what extent exemplar-based and prototype predictors can be applied to predict categorization in natural language concepts. Participants categorized novel tropical foods into two well-known natural language concepts: fruits and vegetables. The results indicate that both the prototype predictors and the exemplar predictors contribute significantly to account for the categorization choices, but that the contribution of the prototype predictor comes from just a limited number of features. Keywords: categorization, exemplar models, prototypes, natural language concepts. Categorization in natural concepts 3 In a series of influential papers on the cognitive representation of semantic concepts, Rosch (e.g., Rosch & Mervis, 1975; Rosch, Simpson & Miller, 1976) suggested that artificial category learning is very relevant for studying the mental representation of natural language concepts. Most contemporary authors of review articles (e.g., Komatsu, 1992; Medin, 1989; Medin & Smith, 1984; Smith & Medin, 1981) still subscribe to the same viewpoint. In a typical category learning experiment (e.g., Medin & Schaffer, 1978; Nosofsky, 1988), participants learn a pair of categories (say, C and D) through presentation of exemplars of both categories. Once the participants have mastered the correct labels for all items in the learning set, they are presented with a new set of items (called the “transfer set”) and they are asked to classify these new items in one of the learned categories. The two categories are supposed to function as each other’s contrast. By manipulating the characteristics of the items in the learning set and in the transfer set, it becomes possible to study the mental representation of the newly learned categories. In many of these studies, characteristics of learning and transfer items have been manipulated to differentiate predictions of (1) abstract prototype models (e.g., Reed, 1972), (2) rule-based models (e.g., Busemeyer & Myung, 1992; Vandierendonck, 1995), and (3) exemplar models (e.g., Medin & Schaffer, 1978; Nosofsky, 1988, 1992). In most comparisons, exemplar models provided the better predictions (e.g., Estes, 1986; Hintzman, 1986; Medin, Altom, Edelson & Freko, 1982; Medin, Altom & Murphy, 1984; Medin & Schaffer, 1978; Medin & Schwanenflugel, 1981; Medin & Smith, 1981). For an overview, see Nosofsky (1992). Recently, however, evidence has been presented that exemplar-based and rule-based representations may both be guiding categorization choices (Erickson & Kruschke, 1998; Palmeri & Nosofsky, 1995). (For other hybrid models that combine Categorization in natural concepts 4 exemplar and feature-based information, see Ashby, Alfonso-Reese, Turken & Waldron, 1998; and Smith & Minda, 1998.) Generalizing the findings from artificial category learning to learning and using natural language concepts is not obvious, because of the complexity of most natural language concepts. For instance, for most natural language concepts, it is not clear which features are important in the categorization decisions, while only a few well-defined features are manipulated in most artificial categories. Nevertheless, in everyday life we often come across situations that resemble the categorization experiments well, because people may be required to make categorization decisions involving natural language concepts, whenever new and unfamiliar stimuli have to be sorted into well-learned categories. For instance, new products bought in the supermarket are categorized as “bottles” or “jars” and are labeled accordingly (Malt & Sloman, 1995). In such cases, people use categories that were learned often a long time ago, mostly in childhood. These categorization situations resemble the transfer phase of category learning experiments. In these situations, the mental representations of these well-known natural language concepts cannot be studied by manipulating characteristics of a learning set of items. In this paper, we explored to what extent exemplar-based and prototype models can be applied to categorization in natural language concepts. An experiment was performed in which different sorts of predictor variables were used to predict categorization of novel food items in well-known natural language concepts (i.e., fruits and vegetables). Exemplar predictors were based on the summed (rated) similarity towards well-known exemplars of these categories, and prototype predictors were based on the extent to which features of both categories applied to the presented items. We investigated (1) whether exemplar-based predictors account better for Categorization in natural concepts 5 categorization decisions than prototype predictors, as is usually found in category learning experiments (Nosofsky, 1992), and (2) whether (following Erickson & Kruschke, 1998, and Palmeri & Nosofsky, 1995), there are individual features that contribute significantly in the prediction, over and above what can be explained by the exemplar-based predictors. The Experiment To construct a categorization situation that was as close to the transfer phase of category learning experiments as possible, we wanted to select two natural language categories that were presumed to function as each other’s contrast category and that are embedded in the same concept hierarchy and at the same hierarchical level. Furthermore, we wanted to choose two concepts for which stimuli could be gathered that were novel to the participants and that could easily be believed to belong to one or the other category. Based on these criteria, the concepts fruit and vegetables were chosen for the categorization task. (In the Method section, we will elaborate on the techniques used to select the contrast categories.) Intuitively, fruits and vegetables seem to function as each other’s contrast category and have the same level of abstraction (residing under a more abstract category consisting of edible natural foods). Furthermore and equally important, there exists a rich variety of exotic food with which our potential subjects are unfamiliar and that can be used as stimuli to be categorized in either one of these concepts. Few attempts have been made to apply exemplar models to natural language concepts. (For rare exceptions, see Heit & Barsalou, 1996, and Storms, De Boeck & Ruts, 2000.) Before describing the details of the experiment, we comment on the difficulties in applying exemplar models to predict category decisions for natural language concepts and on how predictions for categorization choices can be derived Categorization in natural concepts 6 from particular versions of the exemplar theory. In the following, we will concentrate on superordinate level natural language concepts (Rosch, Mervis, Gray, Johnson & Boyes-Braem, 1976), because this is the level of abstraction of the concepts fruits and vegetables used in our experiment. We will also briefly review Hampton’s (1979) procedure to derive prototype predictions for natural language concepts. Exemplar models In the context of natural language concepts, it is very hard to derive predictions from an exemplar model with specific memory traces, because it is impossible to find out how many and which exemplar experiences people have stored for concepts at the abstraction level of, for instance, bird or sports. Assuming that the cognitive system only activates a sample of the exemplars stored in memory, a possible sampling process is related to the instantiation principle (Heit & Barsalou, 1996; Storms, De Boeck & Ruts, 2000; De Wilde, Vanoverberghe, Storms & De Boeck, 1999). According to the instantiation principle, a representation of a category includes detailed information about its diverse range of instances, and people generate instances of a category to base category-related decisions on. For example, mammals are assumed to be judged typical for the category animals to the extent that its generated instantiations (e.g., dog, cats, humans, etc.) are typical animals (Heit & Barsalou, 1996). Also, a particular bird x is assumed to be typical to the extent that it resembles other activated instances of the category birds (Storms, De Boeck & Ruts, 2000). An instantiation process can lead us to predict that, when asked to classify a novel stimulus into one of two categories, people will evaluate the similarity of the presented stimulus to activated exemplars of the two categories. Note that this instantiation principle does not specify at what level the instantiation is really represented. In other words, when mammal is (among others) instantiated with an Categorization in natural concepts 7 instance like dog, this may refer to the level of highly specific memory traces, or to the level of an abstract summary representation of all experiences a person has had with dogs. While Heit and Barsalou's model states that only a single instantiation is activated, we will assume that multiple instantiations can be activated and that the exemplars of the two rival concepts that are activated are the same exemplars that are generated in an exemplar generation task. (See also Storms, De Boeck & Ruts, 2000.) Prototype predictions. The prototype theory assumes that people have abstract summary representations of superordinate natural language concepts directly stored in their mental lexicon, by means of lists of characteristic features (Komatsu, 1992; Rosch & Mervis, 1975; Smith & Medin, 1981). Following the procedure described by Hampton (1979), similarity of a set of items to a prototype of a superordinate concept can be derived by first asking subjects to generate features of the concept, and by summing the applicability frequencies of these features for each of the items (possibly weighting the sum according to some importance criterion). Hampton successfully predicted typicalities and response times of well-known stimuli in superordinate natural language concepts using this procedure, but the same prototype predictor can be used to predict categorization of novel stimuli in superordinate concepts. In the experiment, prototype-based and exemplar-based predictions were compared with each other and combined to predict categorization of novel stimuli in two familiar natural language concepts. The predictive power of individual features (or rules) was also evaluated when combined with the exemplar-based predictor. Method Participants. Fifty-six last year high school students and 40 students of the University of Leuven participated in the experiment. Thirty-six high school students Categorization in natural concepts 8 participated in the reminds-me-of task and the remaining 20 high school students participated in the categorization task. Twenty university students participated in a similarity-rating task, and twenty participated in the feature applicability judgment task. All subjects participated voluntarily. The university students were paid for their participation. Material. The two superordinate natural language concepts used in this experiment were fruits and vegetables. To verify whether these concepts function as each other’s contrast category, twenty research assistants were given the standard question for deriving contrast sets (Frake, 1961; Malt & Johnson, 1992; Rosch & Mervis, 1975). They were asked to imagine that they have heard a description of an object, and that they are trying to guess what the object is. Half of the participants then read that their first guess was a vegetable, the other half of the participants read a fruit as their first guess. They were to imagine that the first guess was incorrect and were asked to think about the most plausible second guess. All ten participants that were asked for the contrast category of fruit gave “vegetable” as their second guess. Nine out of the ten remaining participants gave “fruit” as the contrast category for vegetable. The tenth participant answered “a plant”. These results confirmed our intuition that the two concepts function as each other’s contrast category. A collection of 30 different tropical fruits and vegetables were gathered that were presumed to be novel to our student participants. These fruits and vegetables were purchased in specialized shops that import food from Central African and Southeast Asian Countries. Appendix A lists names and continent of origin of the stimuli. Procedure. The 30 stimuli were presented on plates. All 30 plates were placed in a large room on different tables that were separated from each other by wooden Categorization in natural concepts 9 partitions. Participants in all task groups toured the 30 tables in a fixed direction to complete their task. Their starting table was determined randomly. They were allowed to touch and smell the foods, but they were not allowed to squeeze or taste. The presented stimuli were not cut up. They were first asked whether they knew the presented item, and if so, they were asked to write down the name of the item. After answering this question, the procedure was different for the different task groups. In the categorization task group, participants were asked to categorize each of the items into one of the two categories, fruits or vegetables. They were also asked to indicate how sure they felt about their answer on a ten-point rating scale, ranging from 1 (not sure at all) to 10 (very sure). In the similarity rating task, participants rated the similarity of the 30 presented items towards eight exemplars each (four exemplars of the concept fruits and four exemplars of the concept vegetables)1. These target exemplars for every participant were selected randomly from the eight most frequently generated exemplars of the two concepts (taken from an exemplar generation task described in Storms, De Boeck, Van Mechelen & Ruts, 1996), with the restriction that similarities towards all target exemplars (8 for fruits, 8 for vegetables) were rated by 10 different participants2. (In a study of Storms, De Boeck & Ruts, 2000, the instantiation principle was used to predict typicality ratings and response times for well-known, lexicalized items of eight superordinate natural language concepts, including fruits and vegetables. They found that the predictive value of the instantiation principle increased as a function of the number of “best” exemplars taken into account, but only up to approximately seven.) In the feature applicability judgment task, participants were asked to judge whether each feature of a list of 17 features applied to the presented item, by marking Categorization in natural concepts 10 (for every item) each feature with a “1” or a “0”. These features (presented in Appendix B) were selected based on generation frequency in a feature generation task, carried out by thirty other first year psychology students. The 10 most frequently generated features for both concepts were selected for the feature applicability judgment task, but three features were selected for both fruit and vegetable, resulting in a set of 17 different features. In a last task, the reminds-me-of task, participants answered the question what the presented item reminded them of. The goal of this task was to find out which fruit and vegetable exemplars, besides the most generated exemplars from an exemplargeneration task, might be important for the categorization task. No indication was given as to the level of abstraction of the answers. Thus, all answers were registered, whether they were very general (e.g., “a fruit”), very specific (e.g., “a granny smith apple”) or in between (e.g., “an apple”). The instructions encouraged them to give more than one answer for every item. Results and Discussion The few “yes” answers given by the participants to the question whether they knew the presented item were verified. If the answer was correct (which was very rarely the case), then the similarity rating (for participants in the similarity rating task), the classification decision and certainty rating (for participants in the categorization task), the reminds-me-of answer (for participants of the reminds-me-of task) or the applicability decisions (for participants of the feature applicability judgment task) for the corresponding item were discarded from the data set. Two dependent variables and two predictor variables were constructed. The first dependent variable was the proportion of “fruit” classification decisions (which is, of course, in this forced choice situation, the complement of the proportion of Categorization in natural concepts 11 “vegetable” decisions). It is important to note that the proportion of “fruit” classifications differed considerably over the 30 items that were presented. For five items, a unanimous classification was given by all 20 participants (one as “fruit”, four others as “vegetables”). Furthermore, for 9 of the 30 items less than two thirds of the participants gave the majority classification answer. These results show that participants differed in their classification choices and thus that the task was not trivial. The reliability of the classification decisions was estimated by applying the Spearman-Brown formula to the split-half correlation, after randomly dividing the participant group into two groups of equal size. The reliability estimate of the classification decisions was .92. The second dependent variable was also based on the classification decisions but incorporated the certainty ratings. All the classification decisions were transformed to certainty ratings that were defined on a 20-point rating scale. These certainty ratings ranged from 1 to 20, where 20 referred to a “fruit” classification with a maximum certainty, and where 1 referred to a “vegetable” classification with a maximum certainty. “Vegetable” and “fruit” classifications with certainty ratings of 4 were thus transformed in certainty ratings of 7 and 14, respectively. The reliability of the certainty ratings was again estimated using the procedure described above. An estimation of .98 was obtained. The two exemplar-based predictions (one for fruit and one for vegetable) were derived as follows. The instantiation predictions for fruit and for vegetables were calculated by simply summing the similarity ratings over the eight most frequently generated exemplars of the corresponding category. Different weightings were tried out, based on generation frequencies and rank order information in the exemplar Categorization in natural concepts 12 generation task. Since none of these weightings improved the prediction, only analyses based on the unweighted sum are reported here. To derive the two prototype predictions, Hampton’s (1979) procedure was followed. Two prototype scores were calculated. For every item, the applicability frequency of the item to all the 10 vegetable features were summed and the same was done for all the 10 fruit features. The feature frequencies were weighted based on the generation frequency of the corresponding features. (Several feature weightings were tried out, based on the generation frequency of the features and on independently gathered feature importance ratings. The weighting based on the generation frequencies yielded the best predictions. Therefore, in the remainder of this article, we will only present the results of this weighted prototype measure.) Regression analyses were done, in which the two dependent variables were predicted from the two prototype-based predictors (for fruit and for vegetables) and from the two exemplar predictors, res...
View Full Document

  • Fall '19
  • Psychology, Predictor, Categorization, concept learning, Eleanor Rosch

What students are saying

  • Left Quote Icon

    As a current student on this bumpy collegiate pathway, I stumbled upon Course Hero, where I can find study resources for nearly all my courses, get online help from tutors 24/7, and even share my old projects, papers, and lecture notes with other students.

    Student Picture

    Kiran Temple University Fox School of Business ‘17, Course Hero Intern

  • Left Quote Icon

    I cannot even describe how much Course Hero helped me this summer. It’s truly become something I can always rely on and help me. In the end, I was not only able to survive summer classes, but I was able to thrive thanks to Course Hero.

    Student Picture

    Dana University of Pennsylvania ‘17, Course Hero Intern

  • Left Quote Icon

    The ability to access any university’s resources through Course Hero proved invaluable in my case. I was behind on Tulane coursework and actually used UCLA’s materials to help me move forward and get everything together on time.

    Student Picture

    Jill Tulane University ‘16, Course Hero Intern

Stuck? We have tutors online 24/7 who can help you get unstuck.
A+ icon
Ask Expert Tutors You can ask You can ask You can ask (will expire )
Answers in as fast as 15 minutes