UNCERTAINTY AWARE TEMPORAL RELATIONSHIP MODELLING IN KNOWLEDGE GRAPHS

Abstract

Existing approaches for constructing knowledge graphs (KGs) from text sources effectively capture structural relationships but often struggle to represent the temporal uncertainty inherent in event-based information. Their temporal encoding typically relies on exact timestamp matching, which limits the ability to handle natural temporal variation among events. This thesis introduces Probabilistic Temporal Relationship Modeling (PTRM), a framework that represents events as temporal distributions to capture uncertainty in timing and relationship propagation. PTRM employs distributional embeddings that jointly encode semantic and temporal uncertainty, enabling soft temporal matching through probabilistic overlap computation. Experiments involving 200,000 events from the ICEWS and GDELT datasets demonstrate that PTRM substantially outperforms existing baselines, including temporal knowledge graph embedding methods and classical approaches, while maintaining strong calibration properties essential for decision-critical applications. The results demonstrate that PTRM bridges the gap between rigid timestamp-based methods and the temporal flexibility required for real-world reasoning, achieving state-of-the-art performance in temporal relationship inference under uncertainty.