Reorienting the study of language learning using naturalistic data

Session Type: Paper Symposium

Abstract

Theories of language learning acknowledge that input matters, but researchers’ assumptions about what children see, hear, and do in the real world usually lack ecological validity. Unlike older naturalistic corpora, recent naturalistic audio and video datasets sample more frequently, follow children longitudinally, and/or prioritize first-person perspectives that allow researchers to see the world from a child’s view. The current symposium presents four projects that leverage these exciting datasets to uncover novel aspects of developmental environments that challenge long-held assumptions about the nature of early learning.

Paper 1 applies machine learning to longitudinal headcam videos, showing that category knowledge can be learned from sensory input, challenging the necessity of innate biases to explain conceptual development. Paper 2 presents an analysis of home recordings of one-year-olds, revealing that infants hear words for objects that they touch, and that words for frequently touched objects are learned earlier, suggesting that word learning must be studied as an interactive, multi-modal process. Paper 3 combines an eye-tracking experiment and analyses of home recordings of one-year-olds, which together reveal that many of infants’ earliest-learned words are abstract, with highly variable visual input both within and across children. This challenges learning theories and measurements of word comprehension that focus on imageable words and uniform visual input. Paper 4 analyzes day-long audio recordings to reveal that the temporal dynamics of language input vary greatly across children, words, and contexts, suggesting that temporal structure may explain individual and word-type differences in learning trajectories.

Sub Unit

• Language, Communication

Chair

• Erica Wojcik, Skidmore College

Individual Presentations

• Self-supervised learning through the eyes of a child - Non-Presenting Author: A. Emin Orhan, New York University; Non-Presenting Author: Vaibha V Gupta; Presenting Author: Brenden Lake, New York University

• The Active Child: Object Play leads to Word Learning - Presenting Author: Catalina Suarez-Rivera, New York University; Non-Presenting Author: Emily Linn, New York University; Non-Presenting Author: Catherine S Tamis-Lemonda, New York University

• Moving beyond “nouns in the lab”: Using naturalistic data to understand why infants’ first words include uh-oh and hi - Presenting Author: Kennedy Casey, Princeton University; Non-Presenting Author: Christine Potter, University of Texas at El Paso; Non-Presenting Author: Erica Wojcik, Skidmore College; Non-Presenting Author: Casey Lew-Williams, Princeton University

• Temporal dynamics of words in daylong audio recordings - Presenting Author: Jessica L Montag, University of Illinois at Urbana-Champaign