Proceedings of the Annual Meeting of the Cognitive Science Society

One hallmark of human reasoning is that we can bring to beara diverse web of common-sense knowledge in any situation.The vastness of our knowledge poses a challenge for the prac-tical implementation of reasoning systems as well as for ourcognitive theories – how do people represent their common-sense knowledge? On the one hand, our best models of so-phisticated reasoning are top-down, making use primarily ofsymbolically-encoded knowledge. On the other, much of ourunderstanding of the statistical properties of our environmentmay arise in a bottom-up fashion, for example through asso-ciationist learning mechanisms. Indeed, recent advances in AIhave enabled the development of billion-parameter languagemodels that can scour for patterns in gigabytes of text from theweb, picking up a surprising amount of common-sense knowl-edge along the way—but they fail to learn the structure of co-herent reasoning. We propose combining these approaches, byem- bedding language-model-backed primitives into a state-of-the-art probabilistic programming language (PPL). On twoopen-ended reasoning tasks, we show that our PPL modelswith neural knowledge components characterize the distribu-tion of human responses more accurately than the neural lan-guage models alone, raising interesting questions about howpeople might use language as an interface to common-senseknowledge, and suggesting that building probabilistic modelswith neural language-model components may be a promisingapproach for more human-like AI.