Date of Award

Spring 1-1-2016

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

First Advisor

Eliana Colunga

Second Advisor

Michael Mozer

Third Advisor

Aaron Clauset

Fourth Advisor

Matt Jones

Fifth Advisor

Tammy Sumner

Abstract

Network models of language provide a systematic way of linking a child’s current vocabulary knowledge processes to the structure and connectivity of properties of language which promote future lexical learning. Using network growth models, we explore the relational role of language and the influence of linguistic structure on language learning. Previous research has proposed that language is learned by a process of semantic differentiation that can be modeled through a network process of preferential attachment, with highly connected nodes being learned earliest. This model accounts for high-level lexical network structure and also captures empirical age of acquisition reports. Alternately, language learning may be driven by contextual diversity, or the diverse contexts and meanings of unknown words in the environment. In this thesis, we test these and other ideas by extending these models to acquisition trajectories of individual children, predicting the individual words a child is likely to learn next. We explore how the definition of a graph, the assumed network growth process, and measures of node importance affect our ability to model acquisition. We not only construct a theoretical framework for network models of acquisition but also test the ability of these models to account for learning and development. This work suggests that network models provide a framework for understanding the cognitive and developmental processes of language acquisition.

Neural network models, often called connectionist models, offer another independent approach to modeling learning and development. We focus on associations in a child’s current vocabulary that might be relevant and even facilitatory to the learning process of young children by constructing predictive models. The associative learning framework of our neural network models allow for different types and timescales of learning to be captured. A key idea to data-driven neural network models of acquisition is that there are strong similarities among the way in which children learn, but the differences between children are also predictive. Assuming that there are different types of language learners and that the vocabulary (together with child age) at any time point reflects the type of learner a particular child is, machine learning models can provide a powerful and predictive tool to aid with classification and diagnostics of a child’s learning trajectory. Focusing specifically on using a child’s vocabulary to predict future lexical learning. We explore a variety of representations of a child’s current vocabulary knowledge, including those from a productive vocabulary report as well as representations based on natural language processing algorithms, adult norms, and phonemic content. We find that individual words in a child’s vocabulary are informative in predicting future vocabulary growth using a neural network model. These results additionally suggest the need to consider differences amongst learners. Our best performing model has information not only about a child’s own vocabulary knowledge but also about the normative acquisition trends of words in that child’s vocabulary. These two types of information improve predictive accuracy and suggest potential diagnostic and interventional tools for helping bridge the lexical differences of language delayed children and their age-matched peers.

Share

COinS