Formalizing theories of storage versus computation with Tree-based Grammars

In this paper, we will explore why Tree-based Grammars (e.g. Johnson et al., 2007; Joshi & Schabes, 1997; O’Donnell, 2015; Scha et al., 1999) are useful to formalize and test theories of storage versus composition in language. In particular, we highlight how different learning frameworks for these grammars formalize different predictions about which preconstructed chunks will be stored holistically versus which sequences will be constructed compositionally. These frameworks thus provide a principled way of exploring the space of possibilities between full-decomposition theories (in which only the smallest possible units are stored) and full-storage theories (in which anything and everything is stored).