A computational investigation of the Sapir-Whorf hypothesis: The case of spatial relations

Investigations of the Sapir-Whorf hypothesis often ask whether there is a difference in the non-linguistic behavior of speakers of two languages, generally without modeling the underlying process. Such an approach leaves underexplored the relative contributions of language and universal aspects of cognition, and how those contributions differ across languages. We explore the naming and non-linguistic pile-sorting of spatial scenes across speakers of five languages via a computational model grounded in an influential proposal: that language will affect cognition when non-linguistic information is uncertain. We report two findings. First, native language plays a small but significant role in predicting spatial similarity judgments across languages, consistent with earlier findings. Second, the size of the native-language role varies systematically, such that finer-grained semantic systems appear to shape similarity judgments more than coarser-grained systems do. These findings capture the tradeoff between language-specific and universal forces in cognition, and how that tradeoff varies across languages.