When Facts Expire: Hybrid Approaches for Temporal Validation of Facts in Multiple and Heterogeneous Knowledge Graphs

Jury

• Marieke VAN ERP, Reviewer, Chercheuse (equivalent HDR), KNAW Humanities Cluster 
• Oscar CORCHO,  Reviewer, Professeur, Universidad Politécnica de Madrid 
• Catherine FARON,  Examiner, Professeure, Université Côte d’Azur 
• Wassila OUERDANE, Examiner, Professeure, CentraleSupélec, Université Paris-Saclay
• Thomas GUYET,  Examiner, Chargé de recherche, HDR, Centre Inria de Lyon   

Abstract

This doctoral thesis addresses the challenges of correcting and ensuring the completeness of large-scale Knowledge Graphs (KGs), focusing on the temporal validity of facts. KGs often exhibit inaccuracies and gaps, particularly for dynamic facts. Robust correction mechanisms are crucial, as fact veracity depends on the source, extraction, and validity period, making temporal information indispensable for validating relations such as the residence of a person or his employment.

To overcome these limitations, this research develops new frameworks and resources. A new framework improves entity alignment between KGs through the transfer and verification of precomputed keys. To bridge a gap in direct property and class alignment, KG Nexus, a LOD-Cloud catalog, was created. The thesis also explores symbolic temporal constraints (Allen’s Algebra) to validate the temporal context of facts. Finally, a new deep learning framework is presented for temporal knowledge graph embedding tools, relying also on Allen’s Algebra relations, resolving the need to represent every timestamp.

Collectively, these contributions facilitate more robust and user-friendly data validation in KGs, improving LOD Cloud utilization and cross-validation. This research advances KG correction and temporal reasoning, while also highlighting future needs in information validation within temporal knowledge graphs.