Learning about the meaning of verb-particle constructions from corpora

This paper describes a distributional approach to the semantics of verb particle constructions (e.g., put up, make off). In common with many other varieties of multiword expression, verb-particles vary in the extent to which the component words of the phrase contribute independent meanings. A technique for automatically detecting when and how a component word is making such a contribution would be very useful in the construction of lexicons. Using verb particles as a test case we suggest that a comparison of the lexical contexts in which the phrase and the components occur can provide us with vital information in this regard. Our hypothesis is that the lexical contexts in which a given verb-particle construction occurs across a corpus will be more similar to those of a given component word if that component word is contributing an independent meaning to the phrase. We demonstrate a convincing correlation between contextual similarity and the compositionality judgements of expert and non-expert annotators.