Ontology-Guided Extraction Layer for Knowledge Graph Construction from Heterogeneous Documents

A recent paper on arXiv (2607.28662) details the creation, execution, and empirical enhancement of an extraction layer that transforms a live document stream into a validated knowledge graph that adheres to a formal ontology. This system processes document metadata from Kafka, directing PDF, spreadsheet, Office, and image files through specific handlers, and performs entity and relationship extraction in two stages using a locally hosted Qwen3.5-9B model fine-tuned for the ontology. Its key feature is ontology-guided extraction, which retrieves a relevant portion of a curated ontology in real-time from a graph database via embedding similarity, integrating it into the extraction prompt for type consistency. The paper also tackles issues like type vocabulary fractures, name variations, relationship duplication, and the conflation of distinct individuals with the same name, aiming to enhance consistency in knowledge graph development for diverse document streams.