An Ontology Embedding Approach Based on Multiple Neural Networks

Abstract

In this paper, we present a low-dimensional vector representation method for the concepts and instances of an ontology. The main idea is to transform the ontological entities into digestible data for machine learning and deep learning algorithms that only use digital inputs. The generated vectors will represent the semantics contained in the source ontology. We use the semantic relationships connecting the concepts as a landmark to train expert neural networks using the noise contrastive estimation technique to project them into a vector space specific to this relationship with weightings dependent on their frequency. The resulting vectors are then combined and fed into an autoencoder to generate a denser representation. The generated representation vectors can be used to find the semantically similar ontology entities, allowing creating a semantic network automatically. Thus, semantically similar ontology entities will have relatively close corresponding vector representations in the projection space.