Zipf’s Law in Child Language: Evidence from Peer Conversations

Evidence of Zipf’s law in children’s own speech comes almost entirely from parent–child interactions, and it remains unclear whether children’s word-frequency distributions in peer conversations show the same patterns and how they change with age. We analyzed the speech of 70 English-speaking children aged 34–88 months in peer conversations, drawn from four CHILDES corpora. For each child, Zipf–Mandelbrot distributions were fitted by maximum likelihood to the overall, content-word, and function-word distributions at normalized sample sizes. Each child’s distributions were well approximated by the Zipf–Mandelbrot model but could not be distinguished from a lognormal model; they are therefore best described as heavy-tailed and near-Zipfian. Function words dominated the high-frequency end and had steeper distributions than content words. In the three corpora with overlapping age ranges, older children had lower exponents for the overall, content-word, and function-word distributions; this effect rested mainly on one corpus covering about 3 to 5½ years. In exploratory analyses, the age-related decline in the exponent was most consistently accompanied by an increase in content-word types. These findings extend near-Zipfian distributions and their age-related flattening to children’s peer conversations.