What I'm
working on

Current · ChatterLab, University of Chicago

Capturing multiparty interactions across the developmental trajectory

How does a child break into a conversation? I work on detecting the moment a child enters multiparty talk in naturalistic daylong recordings. The diagnostic is semantic alignment. When speakers are engaged in the same interaction, their language converges in meaning, so alignment that rises above baseline can help us identify multiparty talk. I measure it with BERTScore, using transformer embeddings to compare the meaning of neighbouring turns, within the ALIGN 2.0 framework.

Much of my work is the pipeline that makes this analysis possible. I carry ELAN-annotated recordings into the format the tools expect, preserve the messy realities of child speech such as sung and wordless vocalisations, and run alignment measures over the result.