École d'été | 8 juin 2026 : Trust, Semantics, and Large Language Models par Dean Allemang

École d'été | 8 juin 2026 : Trust, Semantics, and Large Language Models par Dean Allemang

by Dean Allemang

Applied Sciences & Engineering AI — Models & Research UYQArtificial intelligence (AI)UYQFGenerative artificial intelligence
🎙 Dean Allemang 👥 2K 📅 July 9, 2026 ⏱ 62 min 👁 15 📄 expert opinion 🧭 2026-08-15
Available in: English (current) Français

Keywords

trustknowledge graphsLLMSPARQLontology

Summary

Dean Allemang’s talk at the UQAM summer school addresses the challenge of trusting information generated by large language models (LLMs). He begins by noting that while LLMs are useful for tasks like summarizing meeting notes or composing emails, they are not trusted for factual questions in enterprise settings. He identifies four pillars of trust that LLMs lack: provenance, auditability, governance, and fidelity. To address this, he proposes an architecture where an LLM translates a natural language question into a formal query (e.g., SPARQL) against a knowledge graph, rather than directly answering. This approach leverages the LLM’s strength in language translation and the knowledge graph’s deterministic querying. He presents experimental results showing that using OWL ontologies and SPARQL yields three times better accuracy than using DDL and SQL for query generation. He conjectures reasons for this, including the visual similarity between ontology, data, and query, and the coherent history of semantic web standards. He also discusses the importance of describing data well and invites further experiments with other formalisms like Cypher. The talk concludes by emphasizing that this approach provides the necessary provenance, auditability, governance, and fidelity for enterprise trust.

189 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides valuable insights into a practical problem: how to make LLMs trustworthy for enterprise use. The proposed architecture is clear and well-argued, building on the speaker’s extensive experience. The experimental evidence, though limited to one study, is compelling and invites replication. The argumentation is solid, with a logical flow from the problem to the solution, and the speaker acknowledges the limits of his conjectures, distinguishing between measured results and opinionated guesswork.

Scientific Rigor, Source Quality, Title Accuracy

The speaker references his own published paper on the experiment comparing OWL/SPARQL with DDL/SQL, and mentions FIBO ontology. He also cites the work of colleagues and the broader semantic web community. The title accurately reflects the content, and the talk is well-structured. The speaker’s credibility is high, given his background and the practical nature of the talk.

145 words

Title / Content Match

The title accurately reflects the content: the talk focuses on trust in LLMs and how knowledge graphs can serve as accountable sources of truth.

Quality & Reliability

8/10

The speaker is a recognized expert in knowledge graphs and semantic web, with a PhD in AI and industry experience. The talk presents a clear architecture for using LLMs with knowledge graphs to ensure trust, backed by a published experiment. However, some claims are based on personal conjecture and the talk is largely opinionated, though grounded in practical experience.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

  • Paper on OWL/SPARQL vs DDL/SQL experiment — Referenced as the study showing three times better performance with OWL/SPARQL.
  • FIBO (Financial Industry Business Ontology) — Used as an example domain ontology in the talk.

Concurring Sources

  • Knowledge Graphs and LLMs — The talk aligns with recent research on combining knowledge graphs with LLMs for improved accuracy and trust.

Dissenting Sources

  • LLM hallucinations — The talk acknowledges that LLMs can hallucinate, but the proposed architecture mitigates this by using deterministic queries.

Contribution & Novelties

The talk presents a novel approach to building trust in LLMs by using knowledge graphs as a source of truth. The key innovation is the ‘LLM as translator’ paradigm, where the LLM converts natural language questions into formal queries, rather than directly generating answers. This allows for provenance, auditability, governance, and fidelity. The experimental comparison between OWL/SPARQL and DDL/SQL provides quantitative evidence for the benefits of semantic web technologies. The talk also invites the community to replicate and extend the experiments, fostering scientific progress.

Pour aller plus loin :

  • Knowledge Graph — Provides background on knowledge graphs.
  • SPARQL — The query language used in the approach.
  • OWL — The ontology language used.
  • FIBO — The financial ontology mentioned.

118 words

Radar Profile

The radar profile shows high scores in information quantity, quality, and technical level, with a slightly lower but still high score in global reliability. This indicates a technically rich and informative talk with strong credibility, though some parts are based on opinion rather than empirical data.

Reliability 8/10

💬 No comments were provided for analysis.