Do contextual word embeddings represent richly subsective adjectives more diversely than intersective adjectives?

Michael Eric Goodale, Salvador Mascarenhas

Abstract: Distributional approaches to adjectival and nominal semantics capture aspects of meaning that have been elusive to mainstream set-theoretic (extensional or intensional) semantics. Yet, classical distributional methods have failed to find marks of well-established theoretical classifications from mainstream semantics, in particular the difference between intensional adjectives like ‘fake’ and intersective adjectives like ‘French’ (Boleda et al. 2013). We show that contextual word embeddings are sensitive to different kinds of context dependency in the adjectival domain (‘good’ > ‘tall’ > ‘French’). This may demonstrate the systematic distributional differences across the typology of adjectives, and possible evidence for a revision of the typology with respect to intensionals.