LINK-KG: LLM-Driven Coreference-Resolved Knowledge Graphs for Human Smuggling Networks
dipak-meher, Carlotta Domeniconi and Guadalupe Correa-Cabrera, arXiv, 30 October 2025. The first source in this wiki that counts duplicate nodes in an LLM-built graph before and after a fix — the number entity-resolution has been asking for since that page was written.
What it builds
A three-stage coreference pipeline that runs ahead of extraction rather than after it. The piece that does the work is a type-specific Prompt Cache: as the model walks a long document chunk by chunk, the cache carries forward what each pronoun and shortened name resolved to, so chunk nine still knows who “he” is. The narrative handed to the extractor is already disambiguated, and the graph comes out of extraction with fewer aliases to merge. This is the opposite of the sift-kg/gbrain arrangement, where the graph is built dirty and cleaned afterwards.
The measurement
16 US federal and state judicial case documents on human smuggling, filed 1994–2024, pulled through Nexis Uni; 7 of them under 2,500 words and 9 over. LLaMA 3.3 70B at temperature 0, served locally through Ollama on an A100, doing both the coreference and the extraction. Baselines are GraphRAG v0.3.2 (adapted with seven entity types and few-shot examples) and the same group’s earlier core-kg.
Duplicate and noise rates, as percentages of nodes:
| short docs (≤2.5k words) | long docs (>2.5k words) | |||
|---|---|---|---|---|
| dup | noise | dup | noise | |
| GraphRAG | 27.02 | 23.59 | 36.01 | 40.71 |
| CORE-KG | 17.00 | 12.41 | 26.10 | 22.82 |
| LINK-KG | 10.61 | 12.28 | 17.78 | 17.57 |
Averaged, that is the abstract’s headline: node duplication down 45.21%, noisy nodes down 32.22% against the baselines.
The absolute numbers, not the reduction
The relative reduction is the paper’s claim. The absolute column is the one this wiki was missing: unmodified GraphRAG duplicates roughly a third of its nodes on a long document, and the best pipeline here still duplicates about one node in six. So the mechanism pages in this corpus have not been describing a rounding error. graph-construction-quality argued from memgraphrag that extraction is the thing to fix; this is the size of the thing.
It also puts a number on a system the corpus already holds. GraphRAG’s answer to dedup was structural — let community detection and summarization absorb near-duplicates rather than merge them. At 36% duplicate nodes going into that summarization step, absorption is being asked to do a lot.
How the count was made, and why that limits it
Intra-type entity pairs compared with RapidFuzz partial_ratio, pairs scoring ≥75% joined into a
similarity graph, connected components treated as duplicate clusters — then a subject-matter
expert reviews every cluster to drop false positives and add duplicates the string matcher missed.
That works because the nodes are people, places and organizations, where surface form carries the identity. It is the same restriction er-evaluation-framework runs into from the other direction, and it leaves the corpus’s hardest case open: nothing here measures duplication among the proposition-shaped nodes rahulnyk-knowledge-graph produces, because no fuzzy matcher and no annotator can cheaply say whether two propositions are the same claim.
The paper is candid that no annotated ground truth exists for knowledge graphs in this domain, which is why the evaluation rests on fuzzy matching plus expert review rather than on a gold graph. There is no limitations section beyond that. One legal corpus, one model, one domain — treat the percentages as the size of the effect, not as a constant.
Related
entity-resolution · core-kg · graph-construction-quality · graphrag · knowledge-graph · er-evaluation-framework · dipak-meher · memgraphrag