Early word learning through communicative inference

Abstract

How do children learn their first words? Do they do it by gradually accumulating information about the co-occurrence of words and their referents over time, or are words learned via quick social inferences linking what speakers are looking at, pointing to, and talking about? Both of these conceptions of early word learning are supported by empirical data. This thesis presents a computational and theoretical framework for unifying these two different ideas by suggesting that early word learning can best be described as a process of joint inferences about speakers' referential intentions and the meanings of words. Chapter 1 describes previous empirical and computational research on "statistical learning"--the ability of learners to use distributional patterns in their language input to learn about the elements and structure of language-and argues that capturing this abifity requires models of learning that describe inferences over structured representations, not just simple statistics. Chapter 2 argues that social signals of speakers' intentions, even eye-gaze and pointing, are at best noisy markers of reference and that in order to take advantage of these signals fully, learners must integrate information across time. Chapter 3 describes the kinds of inferences that learners can make by assuming that speakers are informative with respect to their intended meaning, introducing and testing a formalization of how Grice's pragmatic maxims can be used for word learning. Chapter 4 presents a model of cross-situational intentional word learning that both learns words and infers speakers' referential intentions from labeled corpus data.