When More Evidence Makes Word Learning Less Suspicious

When More Evidence Makes Word Learning Less Suspicious Gavin W. Jenkins (gavin-jenkins@uiowa.edu) Jodi R. Smith (j odi-r-smith@uiowa.edu ) John P. Spencer (john-spencer@uiowa.edu) Larissa K. Samuelson (larissa-samuelson@uiowa.edu) Department of Psychology and Delta Center, E11 Seashore Hall Iowa City, IA 52240 USA Abstract O ne challenging problem that children overcome in learning new words is recognizing the hierarchical category of a label. For instance, one object could be called a Dalmatian, a dog, or an animal. Xu and Tenenbaum (2007) proposed a Bayesian model to explain how 3.5 to 5-year-olds solve this ambiguity. They emphasized children's appreciation for “suspicious coincidences:” a label applied to three identical toys is interpreted more narrowly than a label applied to one toy. Xu and Tenenbaum did not investigate children’s prior category knowledge, however. We replicated their “suspicious coincidence” effect and measured this knowledge. Unexpectedly, children with more category knowledge appreciated “suspicious coincidences” less. In a second experiment, repeatedly emphasizing novel labels caused all children to stop recognizing the “suspicious coincidence.” These data are inconsistent with the Bayesian account and suggest the phenomenon is influenced by subtler aspects of prior knowledge and by task-specific details. Keywords: Word Learning, Categorization, Bayesian Model Introduction A central issue in the study of word learning is how children acquire names for hierarchically nested categories. “Animal,” “mammal,” “dog,” “Black Labrador,” and “Rover” can all apply to the same referent. This presents a challenge to a young word learner, because when a child hears a label applied to a novel object, the correct interpretation is ambiguous in a hierarchically labeled system: does a novel label “fep” combined with an animal refer to the species, to the breed, or is it a proper name? Additionally, some of the tools children usually use to decipher novel word-object mappings become less helpful in the case of hierarchically nested categories. Mutual exclusivity (Markman, 1989) is counterproductive in cases where two words both refer to the same object, but at different hierarchical levels (e.g., “animal” and a novel word for the same thing, like “Dachshund”). Any child relying on mutual exclusivity would fail to learn more than one hierarchical label for an object at a time. Golinkoff, et al.'s (1992) N3C constraint would be counterproductive for the same reasons: it rejects such overlapping labels by design. To approach this problem, Xu and Tenenbaum (2007) recently suggested a Bayesian approach children might use to succeed at learning names for objects at multiple levels in a hierarchy. Specifically, Xu and Tenenbaum suggest that children recognize so-called “suspicious coincidences” when a label is applied to multiple, distinct exemplars that look very similar. For example, a child might hear the word “fep” applied to a Black Labrador dog. After just one labeling event, the word “fep” is ambiguous. Imagine that a few minutes later, however, the child again hears “fep” applied to a different Black Labrador. Now, the child can use Bayesian inference to suppress some possible interpretations: if “fep” refers to all animals, it would be a “suspicious coincidence” for the first two random examples that the child saw to both be examples of the same breed of dog. For the same reason, the evidence would also be suspicious if “fep” refers to only “dogs.” It would not be suspicious at all, however, to see “fep” applied to two Black Labradors in a row if “fep” meant only “Black Labrador.” Xu and Tenenbaum suggest that children can recognize when a label is applied to a “suspiciously” small subset of the possible objects it could refer to, and use this to infer the label carries a narrow meaning. If the child hears “fep” applied to Black Labradors a third and fourth time, the child's narrow interpretation becomes exponentially stronger, and any other interpretation is de-emphasized (see the “size principle,” introduced on page 252 of Xu & Tenenbaum, 2007). Xu and Tenenbaum (2007) experimentally tested children's ability to infer narrow meanings in “suspicious coincidence” situations. Participants (42-60 months of age) who were shown one stuffed toy Dalmatian labeled “fep” later generalized the label to a variety of toys at different levels of hierarchy: other Dalmatians, different breeds of dogs, and even a few other species of animals, like seals. However, when participants were shown three separate Dalmatians, all labeled “fep,” they almost never generalized the label to anything but other Dalmatians. In this experiment, separate toys were all presented simultaneously, so there was no ambiguity about whether the toys were unique or whether they were the same toy. Xu and Tenenbaum explain this “suspicious coincidence” behavior as a natural extension of a Bayesian model of word learning. The suspicious coincidence is a conceptually important phenomenon because it is not predicted by other models of word learning. Nevertheless, relatively little is known about how this phenomenon is related to the dramatic changes in word learning that take place in early development. Thus, the goal of the present study was to examine how this phenomenon is related to children's emerging category knowledge. Prior knowledge and the similarity structure