Syntactic control for compositional vector space models
Michael Moortgat, Gijs Wijnholds · 2015
Abstract. The framework of Compositional Distributional Semantics unifies vector space models for lexical meanings with a compositional account of how these meanings combine into phrases and larger units. The syntactic engines that have been used to drive the interpretation process (Lambek grammars, pregroups) are problematic in two respects (overgeneration and undergeneration) compromising the accuracy of the quantitative values associated with a derivation. We address these prob-lems by moving to a non-symmetric, non-associative, non-unital type logic with a tree-building tensor operation, generating phrases rather than strings. Composition (tensor) and decomposition of phrases (coten-sor) are treated on a par. Reordering and restructuring are controlled by adjoint pairs of modalities, the grammatical analogues of Linear Logic’s ’!’. We discuss the categorical structures for this model of syntax and the associated graphical language. We identify some empirical areas where the model leads to improved performance.