Directional Context Helps: Guiding Semantic Relatedness Computation by Asymmetric Word Associations

Abstract

Semantic relatedness computation is the task of measuring the degree of relatedness of two concepts. It is a well known problem with applications ranging from computational linguistics to cognitive psychology. In all existing approaches, relatedness is assumed to be symmetric i.e. the relatedness of terms t i and term t j is considered the same as the relatedness of terms t j and t i . However, there are tasks such as free word association, where the association strength assumed to be asymmetric. In free word association, the given term determines the context in which the association strength must be computed. Based on this key observation, the paper presents a new approach to computing term relatedness guided by asymmetric association. The focus of this paper is on using Wikipedia for extracting directional context of each given term and computing the association of input term pair in this context. The proposed approach is generic enough to deal with both symmetric as well as asymmetric relatedness computation problems. Empirical evaluation on multiple benchmark datasets shows encouraging results when our automatically computed relatedness scores are correlated with human judgments.