Abstract

The Role of Difference-Detection in Learning Contrastive Categories Linsey A. Smith ([email protected]) Department of Psychology, 2029 Sheridan Road Evanston, IL 60208 USA Dedre Gentner ([email protected]) Department of Psychology, 2029 Sheridan Road Evanston, IL 60208 USA Abstract Prior research has found that comparison fosters abstraction and transfer of concepts (e.g., categories, solution methods). These learning benefits are often explained by virtue of comparison’s ability to highlight common relational structure between cases. Here we explore the role of comparison in identifying critical differences. Participants compared contrastive cases, listed differences between them, and completed a classification task. We found that carrying out a structural alignment prior to listing differences influenced the kinds of differences people noticed. Further, the kinds of differences people noticed predicted their subsequent classification performance. Keywords: Analogy; structural alignment; comparison; contrast; learning Introduction Comparison has been shown to lead to learning in a number of different realms for both children and adults. Comparing cases facilitates transfer and problem-solving in adults (e.g., Catrambone & Holyoak, 1989; Gentner, Loewenstein, Thompson, & Forbus, 2009; Gick & Holyoak, 1983). Comparison also fosters children’s learning of relational categories (Gentner, Anggoro & Klibanoff, 2011) and relational language (Childers, 2011; Gentner & Namy, 2006; Haryu, Imai & Uchida, 2011). A recent meta-analysis by Alfieri et al., (2013) found that the use of comparison in classrooms is a strong predictor of learning gains. How do these benefits come about? According to structure-mapping theory (Gentner, 1983), when learners compare two cases, they generate a structural alignment between the two representations. This fosters learning in at least three ways (Gentner, 2010; Gentner & Markman, 1997). First, it increases the salience of their common structure; second, it invites inferences from one case to the other; and, third, it highlights alignable differences— differences connected to the common structure. Much of the research showing positive effects of comparison on learning has focused on its effects in abstracting commonalities and inviting inferences (e.g., Catrambone & Holyoak, 1989; Gentner et al, 2009). However, there is mounting evidence that comparison can aid in differentiation as well as in abstraction. For example, comparing two “near-miss” cases (McClure, Friedman, & Forbus, 2010), which are identical except for a crucial structural difference, improves learning (e.g., Gick & Paterson, 1992). Comparison also fosters discrimination between more complex cases, such as alternative solution methods (Rittle-Johnson & Star, 2009), easily confusable concepts (e.g., Day, Goldstone & Hills, 2010; VanderStoep & Seifert, 1993), and category exemplars vs. non-exemplars (Gick & Paterson, 1992; Kok, de Bruin, Robben, & van Merrienboer, 2012; Kurtz & Gentner, 2013). For example, Day et al. (2010) found that having middle-school students contrast positive and negative feedback systems could improve classification of new examples. An open question is how exactly the observed learning effects come about in contrastive case comparisons. These findings underscore the benefit of contrastive cases in learning. Many of these studies utilize pairs that are highly similar except for the crucial difference. Such pairs have two advantages. First, they are ‘self-aligning”—that is, they are extremely easy to align, even for children and novices. Second, once aligned, they have few or no competing alignable differences besides the key intended difference. For example, Kurtz and Gentner (2013) found that people could identify an error in a skeleton faster if they compared it with a highly alignable correct example than if they compared it with the same correct example mirror-reversed (and thus less perceptually alignable). But not all important distinctions can be illustrated with very close ‘near-miss’ pairs. Many important category distinctions involve moderate similarity, with some overlap and many differences. Here we ask what kinds of learning processes best facilitate learning in these more complex cases, in which pairs from different categories are only moderately similar—not so close as to be “self-aligning.” Because structural alignment highlights not only commonalities but also alignable differences, we propose that explicitly encouraging comparison between members of the two categories will facilitate noticing differences and thereby facilitate learning the category distinction. Prior work has shown a relationship between structural alignment and difference-detection (e.g., Gentner & Gunn, 2001; Sagi et al., 2012) and between comparing contrastive cases and transfer (e.g., Day, Goldstone, & Hills, 2010; Rittle-Johnson & Star, 2009); our goal is to clarify the relationships between these phenomena. Thus, the current study examines the connections between structural alignment, difference-detection, and subsequent ability to classify new examples of the two categories. We chose positive and negative feedback systems as our