Training Distributional Matrices for Dutch Transitive Verbs with an Application in Ambiguous Relative Clauses

Abstract

Distributional semantic models represent word meaning as vectors which reflect word distribution in corpora. The field of compositional distributional semantics investigates how these vectors can be composed to represent constituent or sentence meaning. Within this field, the categorial approach trains words that are assigned function types in typelogical grammars, including verbs, as higher-order tensors. This paper implements this approach by training decoupled verb matrices for Dutch transitive verbs and analysing their performance in derivationally ambiguous Dutch relative clauses. In the training of verb matrices, distributional data were partially imported from Tulkens, Emmery & Daelemans (2016) and partially extracted from the Lassy Groot corpus (Van Noord, 2006). Verb matrices were trained using Ridge regression. Analysing the performance of these matrices in relative clauses, it is found that trained matrices are generally sound, but show very little differentiation between subjects and objects. Possible causes and implications of this surprising result are discussed.