Brandon Roy Thesis Defense
Brandon Roy Thesis Defense: "The Birth of a Word"
Date and time:Nov 30 12 3:00pm - 5:00pm
The first two years of a child's life is a period of remarkable development, especially in the area of language learning. From first productive word use around 12 months of age to the emergence of combinatorial speech in the second year of life, how does a child's everyday experience contribute to his or her early language development?
This work builds from the ground up to study the environmental contributions to early word learning, guided by the idea that the naturally emerging activities and social structures of daily life facilitate communication and provide helpful learning constraints. We investigate this idea by shining the light of big data on early word learning through analysis of the largest-ever corpus of one child's everyday experience at home. Through the Human Speechome Project, the home of a family with a young child was outfitted with a custom audio-video recording system, capturing more than 200,000 hours of audio and video of daily life from birth to age three. The annotated subset of this data, the Speechome Corpus, spans the child's 9-24 month age range and contains more than eight million words of transcribed speech, constituting a detailed record of both the child's input and his linguistic development.
Working with a comprehensive, naturalistic dataset opens new avenues for exploration but also requires new approaches to analysis–questions must be operationalized to leverage the full scale of the data. We begin with the basic task of speech transcription, then identify the "word births"–the child's first productive use of each word in his vocabulary. With the resultant vocabulary growth timeline, we examine the environmental contributions to word learning, such as linguistic and prosodic variables in caregiver input speech. But language is not used in a vacuum and is tied to everyday activity, and as such we investigate spatial and activity contexts related to word learning. Activity contexts, such as "mealtime," are identified both manually as well as by probabilistic topic modeling methods that can scale to large datasets. We find that these new nonlinguistic variables can account for much of the variability in when words are learned, and are complementary to the more traditionally studied linguistic measures. Overall, a broad, rich characterization of the early learning environment is both possible and a fruitful direction for studying how children learn language.