A multisource context-dependent semantic distance between concepts

Abstract

A major lack in the existing semantic similarity methods is that no one takes into account the context or the considered domain. However, two concepts similar in one context may appear completely unrelated in another context. In this paper, our first-level approach is context-dependent. We present a new method that computes semantic similarity in taxonomies by considering the context pattern of the text corpus. In addition, since taxonomies and corpora are interesting resources and each one has its strengths and weaknesses, we propose to combine similarity methods in our second-level multi-source approach. The performed experiments showed that our approach outperforms all the existing approaches. © Springer-Verlag Berlin Heidelberg 2007.