Lexical variability and compositionality: Investigating idiomaticity with distributional semantic models

Abstract

In this work we carried out an idiom type identification task on a set of 90 Italian V-NP and V-PP constructions comprising both idioms and non-idioms. Lexical variants were generated from these expressions by replacing their components with semantically related words extracted distributionally and from the Italian section of MultiWordNet. Idiomatic phrases turned out to be less similar to their lexical variants with respect to non-idiomatic ones in distributional semantic spaces. Different variant-based distributional measures of idiomaticity were tested. Our indices proved reliable in identifying also those idioms whose lexical variants are poorly or not at all attested in our corpus.