Aligning Disparate Data Sources Using Large Language Models

Capture d’écran 2026-07-15 à 01.12.23

Soutenue par Nour Elhouda KIRED, à l’Université de Toulouse (IRIT), le 02/12/2025. Dirigée par Olivier Teste. Excellent travail avec des travaux connexes à la thèse de Wissal. Toujours un plaisir de travailler avec Olivier. [theses.fr] [Rapporteur]

Abstract

Schema matching is a key task in data integration but remains challenging when datasets are incomplete or asymmetric, such as schema-only or instance-only tables. This thesis proposes a unified framework that leverages Large Language Models (LLMs) to generate the missing schema or instance information, enabling effective alignment of disparate tabular data sources. The enriched data are embedded into a shared semantic space and matched using two complementary approaches: (i) an unsupervised framework that combines multiple schema-, instance-, type-, and rule-based matchers, and (ii) a supervised framework based on embedding-derived meta-features and machine learning. Experimental evaluation on benchmark and real-world datasets demonstrates that the proposed methods outperform existing approaches, particularly in challenging incomplete-data scenarios. The results highlight the effectiveness of LLM-based context generation in improving schema matching accuracy and robustness, advancing automated data integration for modern heterogeneous data environments.

Laisser un commentaire