MedHal, Ontology-Constrained Generation & SPARQL Query Generalization

Multiple Presentations

Speakers: Fabrice, Gaetan Butault, Zacharie Garnier-Cuchet

Topics: Datasets, Constrained Generation & Legal Mention Detection and Disambiguation

Presentations

1. MedHal: An Evaluation Dataset for Medical Hallucination Detection
2. Ontology-Constrained Generation of Domain-Specific Clinical Summaries
3. How Structured representation can improve for SPARQL query generalization?

Abstract

Translating questions into SPARQL queries enables Knowledge Base querying, but existing datasets are largely template-based, limiting models’ ability to generalize to naturally phrased questions. We introduce frame-semantic approaches that enhance questions with Frame Semantic Role Labeling (FSRL), and release frame-enriched versions of LC-QuAD 1.0, LC-QuAD 2.0, and QALD-10. Experiments with recent large language models show that integrating frame-based representations improves SPARQL generation, especially in scenarios with unseen templates and naturally phrased questions.