Statistical Techniques for Automatically Inferring the Semantics of Verb-Particle Constructions

This paper describes an investigation of some potential features for a statistical approach to inferring the semantics of verb-particle constructions from corpus data. Verb-particles cause particular problems for the computational semantic analysis of language, because their meaning often cannot be derived through the usual compositional methods of analysis. Two novel techniques are presented which promise to provide information about the nature and extent of composition. The first of these measures the extent to which the verb or particle of any given verb-particle may be replaced with a verb or particle of a similar semantic class to form other verb-particles that are attested in the data. The intuition here is that if it reflects systematic patterns in this way then it is more likely that the verb or particle concerned have their simplex meaning. The second technique measures the degree of semantic relatedness between the verb-particle and its component verb. The intuition here is that if a verb-particle is semantically similar to the verb then it is more likely that the verb contributes its simplex meaning. These two features are then combined and used as training data for a classifier using appropriately annotated data.