Logical metonymy resolution in a words-as-cues framework: beyond type-clash

Abstract

Sentences like The author began the book (logical metonymies / aka coercion) combine a verb requiring an event-denoting object (event-subcategorizing verb, begin) and entity-denoting object (book), and involve the understanding of a covert event (reading). Type-clash (between entity-denoting object and event-subcategorizing verb) has been a classical way to account for such phenomena, and together with selectional restriction violations it is widely invoked in linguistic theory. More recent work in psycholinguistics suggests that these phenomena might be better captured via graded notions such as typicality and thematic fit, defined as the typicality of a filler for an argument slot (e.g. how typical it is that books are objects of read, begin, eat, ?). I present two psycholinguistic experiments showing effects of event typicality in logical metonymy interpretation (the baker began the icing - spreading, The child began the icing -> eating). Distributional Semantic Models are computational models of word meaning based on the assumption that words occurring in similar contexts (and in similar argument positions) are also semantically similar, and represent word meaning with corpus-based co-occurrence vectors. Distributional Memory, with its module ECU (Expectation Composition and Update), is a distributional semantic model of the dynamic composition and update of thematic fit and of the graded effect of thematic fit on expectations about upcoming arguments. In particular, ECU can distinguish expectations of upcoming arguments for different typical event scenarios (Baker spreads icing vs. Child eats icing).

The second part of my talk will present first results from computational models such as ECU on logical metonymy, successfully replicating the pattern of results from the psycholinguistic experiments. Both the psycholinguisitic experiments and the similarity-based models suggest an alternative account of logical metonymy interpretation which redefines the binary notion of type-clash in more graded terms, i.e. as thematic fit, and which is compatible with words-as-cues dynamical models of the lexicon.

Alessandra Zarcone

Professor of Language Technologies and Cognitive Assistants

Computational linguist with a background in NLP and in psycholinguistics, working on AI, NLP and human-machine interaction.