Embedding Probabilities in Predication Space with Hermitian Holographic Reduced Representations

Abstract. Predication-based Semantic Indexing (PSI) is an approach to generating high-dimensional vector representations of concept-relation-concept triplets. In this paper, we develop a variant of PSI that accommodates estimation of the probability of encountering a particular predication (such as fluoxetine TREATS major depressive disorder) in a collection of predications concerning a concept of interest (such as major depressive disorder). PSI leverages reversible vector transformations provided by representational approaches known as Vector Symbolic Architectures (VSA). To embed probabilities we develop a novel VSA variant, Hermitian Holographic Reduced Representations, with improvements in predictive modeling experiments. The probabilistic interpretation this facilitates reveals previously unrecognized connections between PSI and quantum theory - perhaps most notably that PSI’s estimation of relatedness across multiple reasoning pathways corresponds to the estimation of the pro...

• Type: Journal
• Year: 2015
• Where: QI
• Authors: Trevor Cohen, Dominic Widdows
• Keywords: Physics, QI 2015