Improving unsupervised vector-space thematic fit evaluation via role-filler prototype clustering