Input Properties Shape Word Segmentation Performance Across Child Development: A Computational Modeling Study

This study investigates how the linguistic properties of child-directed speech (CDS) change across child development and how these changes affect segmentation performance in computational models. We analyzed a cross-sectional corpus of Korean CDS directed at 35 children between 6 and 30 months of age, observing developmental changes in several linguistic properties, including the proportion of one-word utterances, hapax legomena, and monosyllabic words, as well as utterance and word length, lexical diversity (MATTR), entropy, and the use of onomatopoeic and playful words. Simulated word segmentation revealed an inverse U-shaped trajectory in segmentability, with a group-level peak during the early speech stage. This developmental pattern was largely explained by changes in the linguistic properties of the input. These findings highlight the importance of modeling input-learnability dynamics and suggest that shifts in CDS structure systematically shape the learnability of linguistic units during early language development.