A Geometric Method for Detecting Semantic Coercion

In this paper we present state-of-the-art results on the computational classification of semantic type coercion, accomplished using a novel geometric method which is both context-sensitive and generalisable. We show that this method improves accuracy on a SemEval dataset over previous work, and gives promising results on a new more challenging experimental setup involving the same data. In addition to a description of our distributional semantic methodology and the results obtained on an established dataset, we offer an overview of the linguistic phenomenon of coercion and an analysis of the geometric features by which our results are achieved.