Supporting inferences in semantic space: representing words as regions

Semantic space models represent the meaning of a word as a vector in high-dimensional space. They offer a framework in which the meaning representation of a word can be computed from its context, but the question remains how they support inferences. While there has been some work on paraphrase-based inferences in semantic space, it is not clear how semantic space models would support inferences involving hyponymy, like horse ran → animal moved. In this paper, we first discuss what a point in semantic space stands for, contrasting semantic space with Gärdenforsian conceptual space. Building on this, we propose an extension of the semantic space representation from a point to a region. We present a model for learning a region representation for word meaning in semantic space, based on the fact that points at close distance tend to represent similar meanings. We show that this model can be used to predict, with high precision, when a hyponymybased inference rule is applicable. Moving beyond paraphrase-based and hyponymy-based inference rules, we last discuss in what way semantic space models can support inferences.