Semi-Automatic Building of Ontologies from Unstructured French Texts: Industrial Case Study

Abstract An enterprise’s knowledge is often contained in many documents of different types: textual and multimedia. Exploiting this information is generally difficult due to a lack of suitable tools. In this research work, our motivation is to assist non-expert users in creating a semantic model with minimum manual design. This paper describes a semi-automatic pipeline to build an ontology version from French texts by parsing natural language content and exploring open knowledge graphs. Specifically, with practical examples, we showcase how we integrated some existing open-source NLP solutions related to Named Entity Recognition, Relation Recognition and Named Entity Linking for building enriched ontology versions. We applied our approach to generate and manage industrial ontology versions for two real case studies: electrical grid management and requirements engineering. Finally, we conclude that having a complete and enriched knowledge graph requires larger training datasets for better performance of NLP tasks. Some steps of ontology implementation are still hard to computerize completely, notably axiom discovery, which strongly relies on supervised language models.