3-102 - Understanding the mix: Clear cases and noisy data in word-referent learning

Session Type: Paper Symposium

Integrative Statement

Theories of word learning begin with the assumption that the input is noisy: words co-occur with many potential referents. Models and experiments have shown how early word learners might make use of this noisy data through statistical learning. Others have argued the main input for learning are the clear cases when the intended referent –via social or perceptual constraints -- is not ambiguous at all. This symposium brings together researchers on both sides. The first paper –analyzing parent input to children --presents evidence that there are many clear cases and that these predict learning. The second paper presents data and a model suggesting that social cues not only increase attention to the intended referent of a word but limit tracking of other concurrent associations, a finding that fits the hypothesis that it is the clear cases that matter. The third talk presents analyses of how parents name multiple objects over the course of interaction and shows that clear and not clear naming moments are interleaved; the presenters present a model to argue that the dynamic structure of these multiple naming events – clear and unclear -- accounts for learning better than the clear cases alone. The final paper takes a broader look at how seemingly chaotic (win-stay, lose shift) processes over noisy data can yield rational learning. Amidst the different theoretical orientations of the speakers, the symposium converges around the idea that understanding early world learning requires understand the mix of the clear and ambiguous cases.

Sub Unit

• Language, Communication

Chair

• Linda B Smith, Indiana University

Individual Presentations

• The dynamics of reference in the home: Finding the gems among the junk - Presenting Author: John C. Trueswell, University of Pennsylvania; Yi Lin, University of Pennsylvania; Benjamin Armstrong, University of Pennsylvania; Erica Cartmill, University of Chicago; Susan Goldin-Meadow, University of Chicago; Lila Gleitman

• Social cues modulate the strength of encoding alternative referents in cross-situational word learning - Presenting Author: Kyle MacDonald, Stanford University; Daniel Yurovsky, Stanford University; Michael C. Frank, Stanford University

• The learning data from the learner’s perspective - Presenting Author: Linda B Smith, Indiana University; Chen Yu, Indiana University; Sumarga H Suanda, Indiana University; Caitlin Fausey, University of Oregon; Daniel Yurovsky, Stanford University

• Algorithms for statistical learning: The win-stay, lose-sample strategy and preschooler’s causal inferences - Presenting Author: Elizabeth Baraff Bonawitz, Rutgers University - Campus at Newark; Stephanie Denison, University of Waterloo; Alison Gopnik, Univ of California-Berkeley; Thomas L. Griffiths, University of California, Berkeley