Learning Communicative Acts in Children’s Conversations:

A Hidden Topic Markov Model Analysis of the CHILDES Corpus

Voiceover

Presenter(s)

Zoe Marshall

Abstract or Description

We aimed to study communicative skill development in children with a quantitative, natural language processing model. We used a Hidden Topic Markov Model on the CHILDES Corpus, a database of parent and child conversations, in order to capture an aspect of communication that went beyond vocabulary and grammar expansion. We looked at changes in act usage over the course of a single conversation, as well as over development. We characterized speech act diversity and makeup over development, as well as mutual information, which is a measure of speech act predictability, in parent to child, child to parent, and child to child sequences of utterances. We found that children start with just a few communicative acts and quickly branch out as they grow, in patterns that are reflective of previous developmental research. Also, children produce acts contingently, based on previous utterances, and that their contingency increases with age to a point. More research is required to look into what happens after that point is reached. We see this model as a viable alternative to hand-coding speech acts, and hope that this study can increase understanding of the development of these acts in future studies.

Mentor

Dan Yurovsky