This paper investigates the representation of proper names in distributional semantics. We define three properties we expect names to display: uniqueness (being a unique entity), instantiation (being an instance of a relevant kind) and individuality (being separable from the subspace of concepts). We show that taking a standard distribution as the representation of a name does not satisfy those properties particularly well. We propose an alternative method to compute a name vector, which relies on re-weighting the distribution of the appropriate named entity type - in effect, producing an individual out of a kind. We illustrate the behaviour of such representations over some characters from two English novels.
展开▼
