Title:Â From Incremental to Dynamic Temporal Knowledge Graph Construction Using Large Language Models
Summary:Â Most available data is unstructured, making automated knowledge extraction a critical yet challenging task. Knowledge Graphs (KGs) offer a powerful framework for structuring this data, enabling effective information retrieval, inference, and reasoning. However, existing construction approaches face limitations, including semantic entity duplication, dependency on predefined ontologies, and an inability to capture the time-sensitive nature of real-world information.
This talk presents two complementary contributions to LLM-based KG construction. The first, iText2KG, is an incremental, topic-independent method for constructing KGs from unstructured text in a zero-shot setting, without requiring post-processing. It comprises four modules :Â Document Distiller, Incremental Entity Extractor, Incremental Relation Extractor, and Graph Integrator. It demonstrates superior performance across diverse construction scenarios, including scientific papers, websites, and CVs.
The second, ATOMÂ (AdapTive and OptiMized), addresses the limitations of existing approaches, which often suffer from instability across multiple runs, incomplete coverage of key facts, limited scalability, and an inability to capture the time-sensitive nature of real-world data. To address these challenges, ATOM splits input documents into minimal, self-contained atomic facts, improving extraction exhaustivity and stability. It constructs Temporal Knowledge Graphs (TKGs) using a dual-time model that distinguishes between observation time and validity time, and merges the resulting graphs in parallel to ensure scalability. Empirical evaluations demonstrate 18% higher exhaustivity, 33% better stability, and over 90% latency reduction compared to baseline methods.
Français 
English (UK)