Space projections as distributional models for semantic composition

Abstract

Empirical distributional methods account for the meaning of syntactic structures by combining word vectors according to algebraic operators. In this paper, a novel approach for semantic composition based on space projection techniques over lexical vector representations is proposed. In line with the principle of compositionality, the meaning of a phrase is modeled in terms of the subset of properties shared by co-occurring words. Syntactic bi-grams are thus projected in the so called Support Subspace, corresponding to such properties. State-of-the-art results are achieved in a well known phrase similarity task, used as a benchmark for this class of methods. © 2012 Springer-Verlag.