Mapping Lexical Knowledge to Distributed Models for Ontology Concept Invention

Abstract

Ontologies are largely used but the abstraction process required to create them is a complex task that leads to incompleteness. Concept invention offers a valid solution to extending ontologies by creating novel and meaningful concepts starting from previous knowledge. The use of distributed vector representations to encode knowledge has become a popular method in both NLP and Knowledge Representation. In this paper, we show how concept invention can be complemented with distributed representation models to perform ontology completion tasks starting from lexical knowledge. We propose a first approach based on a deep neural network trained over distributed representations of words and ontological concept. With this model, we devise a method to generate distributed representations for novel and unseen concepts and we introduce a methodology to evaluate these representations. Experiments show that, despite some limitations, our model provides a promising method for concept invention.