Usage-based approaches to language acquisition (e.g. Tomasello, 2003) propose that children use multi-word utterances – chunks – to build up grammatical knowledge from recurring patterns in their linguistic input. We investigate the changing influence of this statistical, chunk-based learning on children’s language production over time using the CAPPUCCINO model (McCauley & Christiansen, 2011). This model simulates child language production using chunks extracted from caregivers’ speech.
We selected orthographic transcriptions of conversations between 6 North American children and their caregivers, by sampling transcripts at 6-month intervals between 1;0 and 4;0 (Providence; Demuth, Culbertson, & Alter, 2006). The model parsed caregivers’ utterances for each child by comparing the transitional probabilities between words to a running average transitional probability, making splits between word chunks when the transitional probability between two words dropped below the current average. At the same time, the model also tracked the transitional probabilities between these discovered chunks. After training the model, we simulated children’s sentence production by reconstructing the utterances they actually used in the transcript from the chunk-to-chunk probabilities detected in the caregivers’ speech.
The number of child utterances that were reconstructed correctly based on transitional probabilities between chunks in the caregivers’ speech decreased over time (β = - 0.720, SE = 0.157, p < 0.001). However, the number of child utterances that contained words or chunks the caregivers did not use, increased (β = 0.547, SE = 0.064, p < 0.001). In other words, these results indicate that, over time, children’s speech less directly imitates chunk sequences in their caregivers’ speech, partly because their chunk combinations become more inventive. We discuss how these findings fit within broader usage-based approaches to language acquisition.