Towards Faithful Natural Language Explanations of Ontological Reasoning Using LLMs

Position

Master's or engineering internship at the DISP research laboratory, INSA Lyon, supervised by Linda ELMHADHBI and Sara BOUGUELIA.

• University: INSA Lyon
• Research laboratory: DISP
• Location: Laboratoire DISP, INSA de Lyon, Campus LyonTech La Doua, bât. Léonard de Vinci, 21 av. Jean Capelle, 69621 Villeurbanne cedex, France
• Starting date: February 2027
• Funding: Around €600 per month

Research topic

Ontologies provide a formal way to represent and structure domain knowledge through concepts, properties, relationships, axioms, and logical constraints. Ontology reasoners analyze this knowledge to verify consistency, identify contradictions or unsatisfiable classes, and infer implicit knowledge. Their results may include logical justifications identifying the axioms that contributed to an inference or inconsistency.

These results are generally presented in technical and formal formats, which can be difficult to understand for users who are not familiar with the underlying logical formalisms. The internship will investigate how logical justifications produced by an ontology reasoner can guide a Large Language Model (LLM) in generating natural-language explanations that remain faithful to the reasoning process and do not introduce unsupported information or conclusions.

The work may lead to an initial proof of concept in the form of a plugin integrated into the Protégé environment. The plugin would generate natural-language explanations of ontology reasoning results, including inferred knowledge, inconsistencies, and unsatisfiable classes.

Expected deliverables

• A comprehensive state of the art on ontologies, ontology reasoners, logical justifications, and LLM-based natural-language explanation generation.
• A methodology for generating faithful natural-language explanations of ontology reasoning results, leveraging logical justifications to guide LLMs and prevent unsupported inferences.
• A functional software prototype in the form of a Protégé plugin integrating an LLM to generate natural-language explanations of reasoning results.
• An experimental evaluation on selected ontologies and representative test cases, assessing the faithfulness, comprehensibility, relevance, and accuracy of generated explanations.
• An analysis of the experimental results, including the strengths and limitations of the approach and potential directions for improvement.
• A scientific publication reporting the proposed approach and experimental results.

Prerequisites

• Master's student in Computer Science or equivalent degree.
• Good programming skills, particularly in Python.
• Good knowledge of ontologies and logical reasoning.
• Interest in Large Language Models (LLMs).

Keywords

Ontologies, Knowledge Graphs, Large Language Models (LLMs), Logical Reasoning.

Selection process

Applications will be reviewed through an initial selection process based on the submitted documents. Shortlisted candidates will be invited for an interview.