Thesis (Ph. D.)--University of Rochester. Department of Brain & Cognitive Sciences, Department of Computer Science, 2019.
Chapter 3 was written with Steven Piantadosi and published as Mollica and Piantadosi (2017a) in Open Mind. Chapter 5 was written with Steven Piantadosi and published as Mollica and Piantadosi (2019) in Royal Society Open Science.
This thesis develops the hypothesis that the systematic patterns of children’s word
use over the course of development are the natural consequence of a sophisticated inductive learning mechanism operating with insufficient data. In this thesis, we sketch
out a first-principles account of lexical-conceptual development and implement this
model framework for the case of children learning kinship. Kinship is a valuable
semantic domain to investigate because children show the same developmental trajectory for early word (mis)use, as in their first year of life, spread out over nine
years. A major limitation of evaluating this model and all models of conceptual development is that we have poor intuitions about how children make use of data. To
remedy this, we build a data analysis model to investigate the profile of data usage in
word learning; although this technique will be broadly applicable to developmental
science. We then illustrate how this technique can be used to check the first principles model of inductive learning and investigate the learning process by compiling
a large cross-cultural dataset assessing children’s knowledge of exact number words.
We then take a step back from the learning mechanism and use Fermi-estimation and
information theoretic techniques to quantify the scale of language learning tasks and
highlight the likelihood of sophisticated learning mechanisms for word meanings.