Experimental Results on Exploiting Predicate-Argument Structure for Verb Similarity in Distributional Semantics

An insight from formal semantics is applied to distributional semantics by building verb vectors and tensors that do take into account argument information associated with verbs. Four different argument combination models are presented and used to augment the verb vectors in two different conjunctive and disjunctive ways. The resulting representations are evaluated on a verb similarity task in three different vector spaces. Three different subsets of the similarity dataset are identified and the performance of the models are analysed on them. The overall findings show that the argument-augmented models and in particular a conjunctive model based on point wise multiplication and the Kronecker tensor product performed better than the base line of verb-only vectors and the other operations