Simulating prenatal language exposure in computational models: An exploration study

Area of Science:

• Developmental Psychology
• Computational Linguistics
• Auditory Neuroscience

Background:

• Infant language acquisition is hypothesized to begin prenatally.
• Studies show fetuses and newborns can discriminate native languages.
• Previous computational models of prenatal language exposure (PLE) lacked ecological representativeness.

Purpose of the Study:

• To develop an ecologically representative framework for modeling PLE.
• To simulate language learning with computational models using this framework.
• To compare the effects of PLE versus postnatal-only input on infant language phenomena.

Main Methods:

• Developed a framework modeling prenatal speech input quantity and quality.
• Incorporated empirical estimates of prenatal speech exposure.
• Modeled speech signal attenuation to the fetal auditory system.
• Conducted unsupervised learning simulations with computational models.

Main Results:

• Incorporating PLE affects computational models' language learning outcomes.
• Differences observed between full-term and preterm infant models.
• Duration of PLE influences model behavior depending on the linguistic task.
• PLE inclusion did not enhance model compatibility with empirical infant data.

Conclusions:

• Prenatal language exposure is a relevant factor for computational modeling of infant language acquisition.
• The developed framework provides a basis for future computational studies on the prenatal period.
• Further research is needed to refine models and their alignment with empirical infant data.