Multi-modal learning for construction and enrichment of temporal relations in knowledge graphsKnez, Timotej (Avtor) Žitnik, Slavko (Mentor) temporal relationsinformation extractionlanguage modelsmulti-modalityIn many real-world applications ranging from clinical diagnostics to digital news modeling the ability to understand what happened, when, and in what order is essential. Time is a fundamental dimension in human knowledge and reasoning, and yet most structured knowledge representations fail to adequately capture temporal dynamics. Traditional knowledge graphs predominantly focus on static, entity-centric facts, lacking support for modeling evolving events and their temporal relations. To address this gap, temporal knowledge graphs provide a structured, machine-readable framework for representing events and the temporal, causal, or contextual links between them. This dissertation presents a framework for the construction and enrichment of event-centric temporal knowledge graphs from unstructured textual data, using multi-modal learning approaches. The work is organized around three core contributions. First, it introduces a novel hybrid architecture for temporal relation extraction that combines textual features from large pretrained language models with structured context derived from external knowledge graphs. This integration enables the model to infer nuanced temporal relationships that may not be explicitly stated in the text, particularly benefiting domains where background knowledge is crucial, such as clinical narratives. Second, the dissertation proposes an end-to-end pipeline that unifies event detection, temporal relation classification, coreference resolution, and graph construction. This pipeline is capable of incrementally enriching the knowledge graph as new information becomes available, enabling dynamic and scalable temporal reasoning. The design focuses on medical applications such as electronic health record analysis, where accurate temporal modeling supports decision-making and treatment planning. Third, the work advances the field by tackling fine-grained temporal information extraction. Instead of relying solely on categorical temporal links (e.g., "before", "after"), the framework explores methods for estimating absolute timestamps and event durations using large language models and supervised learning. These techniques allow the construction of precise event timelines, which are valuable in time-critical domains like healthcare, legal analysis, and historical research. Through extensive empirical evaluation on the medical domain, the dissertation demonstrates that incorporating multi-modal knowledge substantially improves the accuracy, robustness, and applicability of temporal relation extraction. The proposed models and methodologies set a foundation for the next generation of temporal reasoning systems and open new possibilities for timeline generation, question answering, and event forecasting in both medical and potentially other domains.20262026-08-06 14:05:05Doktorsko delo/naloga185483VisID: 37213COBISS_ID: 290985987sl