CompCog: Modeling cognitive mechanisms of sound patterns in human language

This project investigates cognitive mechanisms that enable humans to rapidly learn structured language systems. The investigators use behavioral data and computation modeling to test hypotheses about domain-specific and domain-general cognitive processes that underpin efficient human learning of language sound systems. The project addresses broader questions about how humans identify regularities, generalize patterns, and learn structured systems efficiently, advancing understanding about human cognitive abilities and informing the development of more efficient AI systems. This project uses a combination of behavioral experiments, computational modeling, and meta-analytic synthesis of evidence from the published literature to establish a clearer understanding of what, if any, innate preferences human language learners bring to the task of language acquisition. The project emphasizes high-powered replications of existing work to support gold standard science, coupled with novel experimental manipulations that can disentangle the specific mechanisms by which inductive biases guide language learning, and the publication of an open database for centralizing and sustaining evidence in the field. The project leverages speech-trained neural networks to act as an estimate of what is in principle learnable from the speech signal, given only a weakly biased domain-general learner, which can serve as an important comparison point to the results of the experimental work with speakers. Moreover, the project endows speech-trained language models with inductive biases that humans are demonstrated to have, paving the way for more data- and energy-efficient AI models. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.