Individual Submission Summary
Share...

Direct link:

Children Hear More About What is Atypical Than What is Typical

Fri, April 9, 2:45 to 4:15pm EDT (2:45 to 4:15pm EDT), Virtual

Abstract

Language conveys a great deal of information for learning about the world. Statistical models trained on language alone approximate human semantic judgments (Mikolov et al., 2013). Without direct visual experience, children born congenitally blind acquire meanings for visual words (e.g., look) similar to sighted children (Landau, Gleitman, & Landau, 2009). While language must contain information about the world, what information does it contain about what is typical in the world?

Communicative pressures push speakers to note things that surprise us and our conversational partners, rather than to describe what is generally assumed (Gleitman, 1990). Adults tend to remark on what is unexpected given their world knowledge—for instance, noting atypical features of objects (e.g., “the purple carrot”) while allowing typical features to go unnoted (e.g., “the [orange] carrot”) (Rubio-Fernández, 2016). Typical description may be more common in child-directed speech if adults calibrate their speech to align with their child’s nascent conceptual knowledge (Yurovsky, 2018). Nevertheless, we propose that although child-directed speech contains much information about the world, there may be relatively little information about the typical features of things.

We examined parents’ speech to children (ages 14–58 months) in a large longitudinal corpus (Goldin-Meadow et al., 2014), extracting nearly 2000 unique co-occurring concrete adjective–noun combinations from individual utterances (e.g., wooden — shoe from “especially with wooden shoes”). We recruited 444 MTurkers to rate the typicality of these adjective–noun pairs (e.g., “How common is it for [a shoe] to be a [wooden shoe]?”) on a seven-point scale (1—never to 7—always). Human typicality ratings were combined with corpus token frequency to determine the distribution of description over this typicality scale.

As predicted, we found that adjectives in speech to children reliably highlight atypical features of concrete concepts more than typical features (β = -.77, p < .001; Figure 1). We also found that parents’ use of adjectives changes over development: Adjectives directed to younger children describe relatively more typical features of objects (β = .11, p < .001; Figure 2). But at every age, children hear more about atypical than typical features of concrete concepts, and thus simple associative learning from language could lead children to overrepresent atypical features. To see if objects are nevertheless associated with their typical features by second-order word co-occurrence, we trained a model that represents words using their contexts (word2vec: Mikolov et al.,

2013). The learned representation did not reliably represent nouns as more similar to their typical descriptors than their atypical ones. In sum, child-directed speech tends to emphasize atypical over typical features, and typicality is not easily recovered from second-order co-occurrence.

As children start learning about the world, the language environment teaches them more about what is unexpected than expected, at least among concrete concepts. Our work shows that language provides a biased curriculum for learning what is typical—one which misleads simple associative learning mechanisms—and thus children may need to integrate perceptual information or detect other cues in the linguistic signal to learn what is typical in the world.

Authors