Late-Talking Children Talk More? A Machine Learning Approach to Speech Act Analysis in Early Childhood

Abstract

Speech acts shape early language development and social cognition, yet little is known about how late-talking (LT) children use them to achieve communicative goals. We compared LT and typically developing (TD) preschoolers (1;09-6;00) across nine dyadic English corpora, using a Conditional Random Field model to annotate speech acts. We analyzed speech act distributions, hierarchical relations, and contingent responses to assess production and comprehension. TD children produced more declarative statements and wh-questions, whereas LT children produced more unclear word-like utterances and showed reduced comprehension ability (1;09-2;07). Speech acts classified LT and TD groups with 72.3% accuracy, improving to 76.6% with linguistic and demographic features. Classification was driven by co-occurring patterns of speech act frequencies. LT children showed delayed onset of speech acts but employed more speech acts after 3;09, focusing on speaker-centered goals, whereas TD children favored collaborative use of speech acts, revealing complex dynamics in the development of communicative skills.