Obtaining Word Embedding from Existing Classification Model

Martin Sustek, Frantisek Vítezslav Zboril

Abstract: This paper introduces a new technique to inspect relations between classes in a classification model. The method is built on the assumption that it is easier to distinguish some classes than others. The harder the distinction is, the more similar the objects are. Simple application demonstrating this approach was implemented and obtained class representations in a vector space are discussed. Created representation can be treated as word embedding where the words are represented by the classes. As an addition, potential usages and characteristics are discussed including a knowledge base.