Don't Let Perfect be the Enemy of Good: A Comparative Approach to Computational Modeling

Abstract

Scaff et al. present a comprehensive analysis of the CHILDES database, demonstrating clear demographic biases in its naturalistic language recordings. We concur with their conclusion that researchers need to be mindful of these biases when making theoretical claims based on CHILDES data. While we agree that more diversity is needed in future corpus collections, we also argue that a comparative approach to corpus analyses and computational modeling might help alleviate some of the current limitations of CHILDES.

Summary

• Scarff et al. show that there are substantial demographic biases in the CHILDES database.
• We agree with Scarff et al. that these biases need to be considered when theorizing about language development based on CHILDES data.
• We also concur that more support and effort is needed to ensure increased diversity in future collections of naturalistic language recordings.
• However, we contend that we can still learn much from the currently available corpora in CHILDES.
• We argue for a fundamentally comparative approach to using CHILDES data, promising to alleviate some of its current limitations.