Jernej Stegnar, Jože M. Rožanec, Gregor Leban and Dunja Mladenić

ABSTRACT
When building a causal graph from textual sources, such as media
reports, a key task is to provide an accurate semantic understanding of the causal variables encoded as nodes and to link themto existing ontologies with at least two purposes: (i) expand the
knowledge with the domain knowledge captured in such ontologies and (ii) provide accurate and different levels of abstraction of the extracted causal variables. This article describes how we used OntoGPT, a tool for matching raw text to ontology concepts in itially designed for the medical domain, to match concepts from media events to relevant ontologies. We build upon our previous work on extracting causal variables and enrich the extraction pipeline by matching causal variables to concepts from specific domain ontologies. In particular, we describe our work regarding the GEO ontology. Future work will focus on expanding OntoGPT’s capabilities by utilizing a wider selection of ontologies. Addressing its limitations, such as dealing with multiple instances of the same class, will also be crucial for improving its utility. These improvements will allow the tool to better support
strategic foresight applications by providing more detailed insights across a multitude of different sectors, further enriching causal graphs and facilitating more accurate predictive modeling.