Interpolated PLSI for Learning Plausible Verb Arguments

Learning Plausible Verb Arguments allows to automatically learn what kind of activities, where and how, are performed by classes of entities from sparse argument co-occurrences with a verb; this information it is useful for sentence reconstruction tasks. Calvo et al. (2009b) propose a non language-dependent model based on the Word Space Model for calculating the plausibility of candidate arguments given one verb and one argument, and compare with the single latent variable PLSI algorithm method, outperforming it. In this work we replicate their experiments with a different corpus, and explore variants to the PLSI method in order to explore further capabilities of this latter widely used technique. Particularly, we propose using an interpolated PLSI scheme that allows the combination of multiple latent semantic variables, and validate it in a task of identifying the real dependency-pair triple with regard to an artificially created one, obtaining up to 83% recall.