Building Ontologies from Text Corpora

Abstract

This paper presents a novel approach of information extraction for building ontologies covering an extensive range of applications drawn from corpora. Our goal is to propose a method that is independent of domains and based on a distributional analysis of semantic units to bring out all the candidates informative elements (concepts, entities, semantic relations, named entities ...). This method is based on a pipeline of four main stages allowing to refine the extraction information from unstructured text in the form of a suite of decomposable representations (sentences in triplets, "argumental structure"...) until to get a consistent final ontology. We applied the pipeline defined in the context of a repeated sampling of 100 articles randomly drawn from text corpus (`Le Monde' with annual version `2013'). For the evaluation results of the trial implementation of our system, we have achieved a level of accuracy at which was up to 74%. We believe from the results obtained that our methodology is quite generic, and can be easily adapted to any new domain.