There are innumerable ways to categorize our world. Yet, children acquire these categories, and their respective labels, rather quickly and accurately. Children, and adults, rely on several heuristics to guide category creation. First, children are constrained by the whole object assumption; when presented with a novel word, children assume that the new label applies to the entire object, as opposed to a single part or characteristic[3].

Second, is the taxonomic constraint on word learning. Taxonomic relationships are hierarchically based—members are related because they belong to the same superordinate category. For example, a terrier is a type of dog, while dogs, more broadly, are a type of animal. Taxonomic relationships are nested in this way; subordinate categories fall into basic categories which comprise superordinate categories.

Figure 1: Taxonomically organized categories

During word learning, the taxonomic constraint argues that children interpret novel noun labels as belonging to basic category objects. That is, a speaker would assume a new word used to refer to a yorkie means “dog”, as opposed to “terrier” or simply “animal”[4]. These basic level terms are often used more frequently and learned earlier than subordinate or superordinate terms. It is more interesting though, to consider how this bias comes to be.

One hypothesis suggests that the basic level provides enough linguistic information for most contexts[6]. For example, knowing that a yorkie is a dog provides ample information about the object at hand. We know it is (probably) furry and friendly based on our prior experience with all dogs. In contrast, knowing that the dog is specifically a terrier tells us comparatively less. At the other end of the spectrum, 'animal' is far too general.

If learners generally reach for the basic level term, it begs the question of how narrower (or more extensive) meanings are learned for new words. Many definitions are created through experience. Viewing multiple exemplars from distinct categories may encourage generalization to a higher level[5]. For example, if terrier, retriever, and pug are all called dog, it wouldn’t make sense for dog to refer to any single one of these subtypes. Similarly, if a dog, cat, and bear are referred to as 'animals' we assume 'animal' applies to an even higher category.

Learning more specific meanings is arguably more challenging. One potential avenue to a develop a more limited interpretation is co-observation, when a new term is paired with a preexisting basic one, such as 'hunting dog'. Heard like this, the learner can assume that a hunting dog is a special case of dog[1]. Recent work by Xu and Tenebaum, though, suggests that statistical tracking may be the driving force in acquiring these relations. When shown a single image of a category, both children and adults exhibit the basic level bias and extend the label accordingly. However, after witnessing three different objects belonging to the same subordinate category, participants were much more likely to only generalize to the smaller group [7]. This phenomenon is known as suspicious coincidence; learners assume that in ostensive contexts exemplars are chosen purposely to be presentative of the intended category. The "coincidence" is resolved by deciding that the label only refers to members of a reduced category.

In a recent paper, Emberson et al. explore a different account of subordinate category learning by focusing on the content of single exemplar. Specifically, they sought to determine if atypical exemplars encourage more specific interpretations. Atypical exemplars are highly salient members of a basic level. They often have specific or memorable properties that distinguish them from other category members. For example, a chihuahua is an atypical example of dog. Consequently, the authors predicted that “a novel label for an atypical exemplar should be more likely to be interpreted as referring to a more narrow taxonomic level”[2].

Figure 2: Typical and Atypical exemplars (Emberson et al., 2019)

Both children (between the ages of four and five) and adults were shown one of three exemplars, either atypical or typical, referred to by a novel label, 'fep'. Afterwards, participants were show an array of eight objects and asked to find any other feps. The array always contained two subordinate matches, two basic matches, and four distractors. The test phase served to determine if the participants were successfully including the atypical exemplars in the intended category.

Figure 3: Sample Array (Emberson et al., 2019)

When participants viewed an unusual object labeled with a new word, they were more likely to apply this label to other unusual objects, as opposed to the more general grouping. These results suggest that the basic-level bias only occurs when an exemplar is typical. Instead, when atypical exemplars are provided, novel labels are more likely to be interpreted narrowly, at the subordinate level. Emberson et al. also replicated the ‘suspicious confidence effect’; witnessing a new label applied to many examples of the same, smaller subordinate category encouraged limited interpretations. These observations, titled the blowfish effect, are likely the result of statistical inference. Learners reason that an atypical exemplar is unlikely to belong to a larger category featuring more typical exemplars [2].

Emberson et al’s work provides new evidence that typicality serves as a source of information for learners making category generalizations. Their results contrast with previous scholarship emphasizing that words are generally assumed to belong to the basic-level. These findings establish typicality as a key player in constructing narrower categories.

References:

  1. Clark, E. V., Gelman, S. A., & Lane, N. M. (1985). Compound Nouns and Category Structure in Young Children. Child Development, 56(1), 84. doi: 10.2307/1130176
  2. Emberson, L. L., Loncar, N., Mazzei, C., Treves, I., & Goldberg, A. E. (2019). The blowfish effect: children and adults use atypical exemplars to infer more narrow categories during word learning. Journal of Child Language, 46(05), 938–954. doi: 10.1017/s0305000919000266
  3. Hollich, G., Golinkoff, R. M., & Hirsh-Pasek, K. (2007). Young children associate novel words with complex objects rather than salient parts. Developmental Psychology, 43(5), 1051–1061. doi: 10.1037/0012-1649.43.5.1051
  4. Markman, E. M. (2002). Categorization and naming in children problems of induction. Cambridge, Mass.: Mit Press.
  5. Jenkins, G. W., Samuelson, L. K., Smith, J. R., & Spencer, J. P. (2015). Non‐Bayesian noun generalization in 3‐ to 5‐year‐old children: probing the role of prior knowledge in the suspicious coincidence effect.Cognitive Science, 39(2), 268–306.
  6. Murphy, G. L., & Brownell, H. H. (1985). Category differentiation in object recognition: Typicality constraints on the basic category advantage. Journal of Experimental Psychology: Learning, Memory, and Cognition, 11(1), 70–84. doi: 10.1037/0278-7393.11.1.70
  7. Xu, F., & Tenenbaum, J. B. (2007). Word learning as Bayesian inference. Psychological Review, 114(2), 245–72.