Applicative structure in vector space models

Márton Makrai, Dávid Márk Nemeskey, András Kornai

Abstract: We introduce a new 50-dimensional embedding obtained by spectral clustering of a graph describing the conceptual structure of the lexicon. We use the embedding directly to investigate sets of antonymic pairs, and indirectly to argue that function application in CVSMs requires not just vectors but two transformations (corresponding to subject and object) as well.