A Combined Approach for Ontology Enrichment from Textual and Open Data

Abstract

This paper proposes an approach for ontology enrichment for automatically labeling documents describing entities, with very specific concepts reflecting specific users’ needs. The peculiarity of this approach is that it addresses a triple challenge: (1) the concepts used for labeling have no direct terminology in the documents, (2) their formal definitions are not initially known, (3) the information useful to label the documents is not necessarily mentioned in them. To solve those problems, we propose to use an existing ontology of the domain of concern and to enrich it with the definitions of the concepts used for labeling. To construct these definitions, we work on a set of manually labeled documents, used as examples. The ontology is populated with information extracted from these documents, and with information coming from external resources (Linked Open Data). The definitions that we want to get can then be learned based on this populated ontology and on the set of labeled documents. Learned definitions are then added to the ontology (ontology enrichment). Hence, whenever new documents of the same domain have to be labeled, the ontology can be populated in the same way and definitions apply, allowing the new documents to be labeled. This approach, named Saupodoc, is a novel approach to ontology population and enrichment, exploiting the foundations of the Semantic Web by combining contributions of text analysis, linked open data extraction, machine learning and reasoning tools. An evaluation, on two application domains, provides quality results and demonstrates the interest of the approach.