Lexical inference as a spatial reasoning problem

Humans are often able to draw plausible conclusions from incomplete information based on background knowledge about the world. As a substantial part of this background knowledge is of a lexical nature, a crucial challenge in automating plausible reasoning consists in learning how different words are semantically related. This paper argues that most of the lexical relations that we need for plausible reasoning can be identified with qualitative spatial relations in semantic spaces, i.e. highdimensional Euclidean spaces in which words are represented as geometric objects. This leads us to treat lexical inference as a qualitative spatial reasoning problem, and allows us to combine distributional representations with relation extraction methods and existing lexical resources.