Researchers build ontology-guided knowledge extraction layer

A team of computer science researchers has published a paper detailing a production-ready extraction layer designed to transform live document streams into high-fidelity knowledge graphs. The system, which processes metadata from Kafka and routes content through format-specific handlers for PDFs, spreadsheets, Office files, and images, uses a locally deployed Qwen3.5-9B model fine-tuned on a formal ontology. Its core innovation lies in ontology-guided extraction: a live slice of a curated ontology is retrieved from a graph database via embedding similarity and injected into the extraction prompt, slashing catalog overhead by 94% compared to static domain slices.

The extracted results undergo a five-stage refinement pipeline that includes deterministic cleaning, cross-chunk merging, a second pass for relationships, six model-free deduplication algorithms, and an embedding resolution subsystem with a conflict guard. In evaluations on intelligence corpora, the system boosted search recall from 70% to 95% while eliminating all false merges and correcting seven classes of quality defects—ranging from single-character text truncations to systematic duplication of entities with title prefixes.