Séminaire de Recherche en Linguistique

Mainstream approaches to Greenberg's Universal 20 are focused on categorically ruling out noun phrase orders which are unattested in the cross-linguistic record. This comes at the expense of explaining dramatic variation among attested patterns, and potentially overfits the typological data. I propose an alternative account which explains Universal 20 as the result of three probabilistic biases: for transparency (isomorphic mappings between meaning and linearization), simplicity (consistent head ordering), and post-nominal adjectives. I provide evidence from a series of experiments that these biases are at work when noun phrase order is learned and improvised in laboratory tasks.