Abstract

In this talk I’ll discuss a novel approach to the established distributional semantic paradigm for modelling words as vectors in semantically productive spaces. Motivated by theoretical insight into the way that word meaning emerges situationally, this methodology seeks to generate ad hoc lexical representations based on context specific input. I will describe a technique for projecting lower dimensional semantic subspaces based on the idea that semantic contexts can be construed as perspectives on a very high dimensional space of co-occurrence statistics. The expectation is that the geometry of these subspaces will correspond to properties that are specific to the conceptual context with which they are associated. I’ll demonstrate how this methodology can be effectively applied to NLP tasks such as analogy completion and word similarity ranking, and then will consider how the methodology might motivate further exploration of various features of distributional semantic models.