Large Language Models (LLMs) are well known for their remarkably fluent natural-language interaction with information systems, but their outputs are not, by themselves, trustworthy: they lack governance, provenance, auditability, and accountability. In this session, I explore an alternative architectural role for LLMs—not as autonomous answer generators, but as semantic interfaces to curated and governed data resources.

I begin by introducing the idea of formally defined, distributed semantics using established standards from the Semantic Web, notably OWL (for defining shared vocabularies and relationships) and SPARQL (for querying data expressed using those vocabularies). As recently as two years ago, the competence demonstrated by LLMs in computer languages came as a surprise. Today, it is no longer surprising that modern LLMs also show strong competence in understanding the languages of the Semantic Web, reflecting their training on widely adopted technical standards.

Building on this observation, I present an architecture in which an LLM translates a natural-language question into a formal query that is validated against an ontology and executed over curated, federated data sources. In this setting, trust derives not from the language model itself, but from the governed resources it accesses and the explicit semantics used to mediate that access. The session concludes with implications for explainability, governance, and the design of trustworthy AI systems.

Références

It's obvious, but is it true? : https://medium.com/@dallemang/its-obvious-but-is-it-true-50013e3ca4d6

Sequeda, J., Allemang, D.T., & Jacob, B. (2023). A Benchmark to Understand the Role of Knowledge Graphs on Large Language Model's Accuracy for Question Answering on Enterprise SQL Databases. Proceedings of the 7th Joint Workshop on Graph Data Management Experiences & Systems (GRADES) and Network Data Analytics (NDA). https://arxiv.org/abs/2311.07509

Allemang, D.T., & Sequeda, J. (2024). Increasing the LLM Accuracy for Question Answering: Ontologies to the Rescue! ArXiv, abs/2405.11706. https://arxiv.org/abs/2405.11706

Allemang, D., Hendler, J., & Gandon, F. (2020). Semantic Web for the working ontologist: Effective modeling for linked data, RDFS, and OWL (3rd ed.). Morgan & Claypool Publishers. ISBN: 978-1-4503-7615-0