Modeling the Influence of Language Input Statistics on Children's Speech Production

Area of Science:

• Child language acquisition
• Computational linguistics
• Developmental psychology

Background:

• Statistical learning is a key mechanism in early language development.
• Backward transitional probability is a proposed statistical learning mechanism.
• Children's speech production shifts from rote memorization to abstract rule use over time.

Purpose of the Study:

• To test if backward transitional probability accurately predicts children's speech productions throughout early development.
• To investigate whether a computational model's predictive accuracy decreases as children age.
• To compare model performance using recently encountered versus cumulative speech input.

Main Methods:

• Trained a Chunk-Based Learner (CBL) computational model on a longitudinal corpus of child-caregiver interactions.
• Assessed the CBL's accuracy in reconstructing children's speech productions.
• Controlled for utterance length and the presence of duplicate chunks in the input data.

Main Results:

• The CBL demonstrated stable accuracy in predicting children's speech productions across the observed developmental period.
• No evidence was found that the model's predictive accuracy decreased with the child's age.
• Both recently encountered and cumulative speech input yielded similar model performance.

Conclusions:

• Backward transitional probability remains a robust predictor of children's speech productions throughout early development.
• The shift from "chunk"-based to abstracted speech production may not significantly impact statistical learning mechanisms as predicted.
• Computational models can effectively capture statistical regularities in child language development over time.