How Children Form Overhypotheses About Animals

• In Event: 3-079 - Boundaries for the biases in reasoning about categories

Integrative Statement

How do children learn about categories from experience with single exemplars of the kind? Consider a child’s first encounter with a squirrel that is eating nuts and sitting on newspaper. When the child encounters another squirrel, s/he is faced with a problem of induction (Hume, 1748/2007) — should s/he draw inferences about what squirrels eat or where they sit? Overhypotheses, or higher-order generalizations (e.g., animal kinds have characteristic diets), may help solve the problem by limiting the lower-level hypotheses that are considered (Goodman, 1955/1983; Quine, 1960). From early in development, simple overhypotheses can be learned (e.g., Dewar & Xu, 2010; Smith et al., 2002). It is an open question, however, whether more complex overhypotheses can be learned and the nature of required input (cf. Shipley, 2000).

In two studies, we explored the formation of overhypotheses about the characteristic properties of different animal kinds in 4- to 6-year-olds. In the teaching phase of Study 1, participants saw blocks of novel properties of three familiar animal kinds. Between participants, we manipulated whether we described the properties using generic or specific statements (e.g., “Dogs…” vs. “This dog…”). We also manipulated whether comparison was promoted by describing either the same type of property (e.g., “…yippets in their eyes”, “…tomas in their eyes”, etc.) or different types of properties (e.g., “…yippets in their eyes”, “…tomas under their skin”, etc.) for each animal. In the test phase, we examined children’s inductive generalizations about unfamiliar animal kinds. If overhypotheses were formed (e.g., animal kinds have particular parts in their eyes), participants should extend newly learned properties from one member to another member of the kind. Children (n = 108) were more likely to generalize properties when they saw exemplars of the same type of property than different types of properties,
F (1,104) = 13.00, p < .001 (see Figure 1). In contrast, generalizations did not depend on the statements used, suggesting that examples that promote comparison (and allow structural alignment) facilitate children’s formation of overhypotheses.

In Study 2, we manipulated only whether comparison was promoted in the teaching phase, and examined children’s questions about unfamiliar animal kinds in the test phase. If overhypotheses were formed, participants should ask questions about the properties in the teaching phase (i.e., physical features, such as eyes). Children (n = 54) asked equal numbers of physical feature questions, but more questions about existing overhypotheses (e.g., diet), when they saw examples of the same type of property and different types of properties, t (52) = -1.71, p = .09; t (52) = 2.16, p = .04, respectively (see Figure 2). This data suggest that the input results in children using existing (but not new) overhypotheses to guide their questions.

Together, the results provide evidence that children form novel overhypotheses about the characteristic properties of animal kinds. However, the scope of this learning was limited – newly formed overhypotheses guided induction, but did not structure children’s acquisition of knowledge. This research contributes to the growing literature that children are rational constructivist learners.

Authors

• Kristan Marchak, University of Michigan
  Presenting Author

• Susan Gelman, University of Michigan
  Non-Presenting Author