Anceresi, Gatti, Vecchi, Marelli and Rinaldi – A MAP OF WORDS 2023

Description

This study tests whether distributional-semantic models (DSMs) can recover medium-scale spatial layouts—specifically underground-station networks—solely from language. Using fastText vectors for station names in five European cities (Berlin, London, Madrid, Milan, Paris), Experiment 1 found positive relationships between linguistic and geographical pairwise distances in four cities (Madrid nonsignificant). Experiment 2 derived linguistic latitude/longitude from similarities to cardinal-direction words and showed these linguistic coordinates predict real geographic coordinates. Experiment 3 focused on London: schematic map distances and coordinates closely matched geographic ones, and language-derived distances aligned equally with geographic and schematic distances. Overall, results indicate that both structural (relative distances) and Euclidean (absolute coordinates) properties of medium-scale maps are encoded in language, supporting the role of domain-general associative learning in forming cognitive maps. Accuracy is lower than for large-scale maps—likely due to polysemy, multiword names, and corpus coverage—implying reconstructability depends on how well space is encoded in linguistic data.