Finding Schizophrenia's Prozac Emergent Relational Similarity in Predication Space

In this paper, we investigate the ability of the Predication-based Semantic Indexing (PSI) approach, which incorporates both symbolic and distributional information, to support inference on the basis of structural similarity. For example, given a pair of related concepts prozac:depression, we attempt to identify concepts that relate to a third concept, such as schizophrenia in the same way. A novel PSI implementation based on Kanerva’s Binary Spatter Code is developed, and evaluated on over 100,000 searches across 180,285 unique concepts and multiple typed relations. PSI is shown to retrieve with accuracy concepts on the basis of shared single and paired relations, given either a single strong example pair, or the superposition of a set of weaker examples. Search space size is identical for single and double relations, providing an efficient means to direct search across predicate paths for the purpose of literature-based discovery.