Skewed distributions facilitate infants' word segmentation

Abstract

Infants can use statistical patterns to segment continuous speech into words, a crucial task in language acquisition. Experimental studies typically investigate this ability using artificial languages with a uniform frequency distribution, where all words occur equally often. However, words in natural language follow a highly skewed distribution conforming to a Zipfian power law, in which few words occur frequently while many occur infrequently. Prior work shows that such skewed distributions facilitate word segmentation, but the experimental evidence for this has been limited to individuals aged ten years or older, leaving unclear whether this effect arises from accumulated linguistic experience or is already present in the early stages of language learning. To address this, we conducted a word segmentation study with 7- to 9-month-old infants. Infants were exposed to a continuous speech stream containing four artificial words, presented either in a uniform or skewed frequency distribution. We found that infants exposed to the skewed distribution showed a greater looking time difference between familiar and unfamiliar words compared to those in the uniform condition. These findings suggest that skewed distributions facilitate learning during early linguistic development, highlighting the impact of such distributions on language acquisition. Moreover, these findings suggest that the widespread use of uniform distributions in lab-based studies may underestimate infants' segmentation abilities.