CORE-KG: An LLM-Driven Knowledge Graph Construction Framework for Human Smuggling Networks
dipak-meher, Carlotta Domeniconi and Guadalupe Correa-Cabrera, arXiv, 20 June 2025. The earlier half of the pair that gives this wiki its first before/after duplicate counts; link-kg is the successor by the same group and supersedes its numbers.
The design
Two steps in sequence: type-aware coreference resolution, then entity and relationship extraction under domain-guided instructions. The diagnosis is that an LLM turned loose on a legal document produces two distinct defects — hallucinated or irrelevant nodes (“legal noise”: docket numbers, statutes, procedural boilerplate) and duplicate nodes from unguided extraction — and that they need separate fixes. Naming them separately is why the evaluation reports two numbers instead of one quality score.
The numbers
Against a GraphRAG baseline on the same judicial-case corpus: 33.28% fewer duplicate nodes and 38.37% less legal noise.
These hold up under independent re-measurement, which is rarer than it should be. link-kg‘s tables re-run CORE-KG as a baseline and land on 17.00% duplicates against GraphRAG’s 27.02% on short documents and 26.10% against 36.01% on long ones — a 37% and a 27.5% relative reduction, averaging close to the 33.28% claimed here. The successor paper had every incentive to report its predecessor unflatteringly and did not.
Why it is filed here
On its own this would be a domain application. What it contributes to this wiki is the baseline row: it is the first of the two sources establishing that unmodified GraphRAG duplicates a large minority of the nodes it extracts, which is what entity-resolution and graph-construction-quality had been arguing without evidence.
Related
link-kg · entity-resolution · graph-construction-quality · graphrag · knowledge-graph · dipak-meher