Preface for the Fifth International Workshop on LLM-integrated Knowledge Graph Generation from Text

Knowledge graph construction from text has long been framed as a pipeline of entity recognition, entity linking, and relation extraction. Recent progress in large language models has radically simplified the production of plausible triples, and systems such as Microsoft GraphRAG have popularized the use of automatically generated graphs for retrieval-augmented question answering. Yet the central challenge remains: how do we move from plausible text-derived triples to trustworthy, reusable, ontology-compliant knowledge graphs? In this talk, the Text2KG task is revisited from both Natural Language Processing (NLP) and Semantic Web perspectives. Open information extraction, which supports exploratory graph construction, is contrasted with ontology-guided information extraction, where extracted knowledge must satisfy explicit conceptual models, domain and range constraints, provenance requirements, and graph-level consistency. Neural and generative approaches to relation extraction, including BERT-based models and REBEL-like sequence-to-sequence models, are discussed, highlighting the need for richer extraction paradigms involving causal relations, events, frames, and n-ary structures for constructing high-quality domain knowledge graphs. The talk presents several case studies, including ontology-based extraction of olfactory heritage knowledge in ODEUROPA, multimodal knowledge graph population in SILKNOW, and recent work on Text2KGBench, its re-annotation, and the development of the PERSEUS model in collaboration with Lettria. These examples demonstrate both the promise and the limitations of current LLM-based approaches, emphasizing challenges related to faithfulness, schema conformance, evaluation, and reproducibility. The talk concludes by arguing that the next generation of Text2KG systems should be evidence-grounded, ontology-aware, multimodal, and benchmark-driven, and introduces the TRIPLET challenge and its joint text–table–knowledge graph information extraction task as an important step toward evaluating knowledge extraction in heterogeneous real-world documents.