The Mapping Global China knowledge graph (data)
This folder ships the finished knowledge graph as a single compressed SQLite
file, mgc_kg.sqlite.gz (about 21 MB). The copilot notebook
(../notebooks/kg_copilot.ipynb) decompresses it once on first run to
mgc_kg.sqlite (about 130 MB) and reads it. You do not need to run the
pipeline or hold any API keys to use it.
Everything here was produced by the pipeline in ../mgc_kg/. Every node,
relationship, and claim records where it came from: a source document, a
link, and the exact character span of the evidence.
How big is it
Built from 5,622 source documents: 5,601 government and company press releases (many originally in Chinese, dates 2022-01-04 to 2026-07-30) and 21 academic papers about China's global engagement (dates 2018 to 2025).
- 44,781 nodes
- 118,205 relationships (edges)
- 22,341 claims
- 4,666 events
- 94 countries
Nodes by type
| Type | Count | What it is |
|---|---|---|
| Claim | 22,341 | An assertion made in a document (a press statement or a paper argument) |
| Project | 5,808 | A concrete project: port, railway, power plant, corridor, zone |
| Document | 5,622 | A source press release or academic paper |
| Event | 4,666 | A dated happening with participants and a location |
| Organization | 3,691 | A company, ministry, bank, or other body |
| Person | 2,166 | A named individual, such as officials and executives |
| Framework | 197 | An initiative or agreement, for example the Belt and Road Initiative |
| Impact | 155 | An environmental, social, or security consequence |
| Country | 94 | A country, keyed by ISO3 code |
| Tag | 32 | A topic tag from academic metadata |
| Sector | 8 | A sector such as transport, energy, or ports and maritime |
| Theme | 1 | A cluster of thematically related papers |
Edges by tier
Edges are grouped by how much to trust them:
| Tier | Count | Meaning |
|---|---|---|
| structural | 25,204 | Deterministic facts from curated datasets and document structure, no guessing |
| inferred | 64,027 | Extracted or computed: relationships, claims, events, co-mentions, similarities |
| provenance | 28,974 | The evidence layer linking documents to the entities and claims they carry |
Most common edge types: MENTIONS (27,595), ABOUT (27,321, a claim to the entities it concerns), MAKES_CLAIM (22,341), HAS_PARTICIPANT (6,455), ISSUED_BY (5,601), LOCATED_IN (5,483), IN_SECTOR (5,317), HAS_EVENT (4,676), IMPLEMENTED_BY (2,683), PART_OF (2,591), INVESTS_IN (1,917, carrying dollar amounts where the dataset provides them), CO_MENTIONS (1,521), SAME_AS (1,379, including cross-language matches), OCCURRED_IN (1,260).
Claims
22,341 claims, each tagged with a stance: describes (10,234), supports (7,757), critiques (4,319), complicates (31). By register: 22,237 are assertive official statements from press releases, and 104 are hedged academic arguments from papers.
What is in the file (schema)
The shipped database contains the tables the copilot uses:
kg_nodes: the entities and claims, one row per node, with a type and a canonical label.kg_edges: the relationships, with a type, a tier, a confidence score, the evidence text, and the source document.kg_claims: the reified claims, with stance, register, evidence span, and the entities each is about.kg_documents: the source documents, including the normalized full text.kg_aliases,kg_pipeline_runs: supporting metadata.
To keep the download small, tables the copilot does not read were removed from the shipped copy: vector embeddings, the pre-resolution staging mentions, and the separate mention and place-mention tables. The full database (about 238 MB) is regenerated by running the pipeline and is not committed.
Provenance
Every relationship and claim carries its source: source_doc_id, a document
URL and title, the character offsets of the evidence inside the document text,
the evidence text itself, and the extractor and model that produced it. Nothing
in the graph is a black box; you can always trace a fact back to the sentence
it came from.
Scope and limitations
- The press corpus is a sample: up to about 200 releases per country across 94 countries, not the entire archive.
- The academic corpus is small (21 papers), and only one had full text available under open access; the rest are abstract-only, which limits how much the paper-side extraction could capture.
- Some of the most newsworthy relationship types (investments, state visits, most impacts) come from the project's curated structured datasets and carry no press-release date, which is why the weekly-headlines view in the notebook widens its window by document count rather than by calendar days.
How it was built
See the top-level ../README.md and the overview at the top of
../notebooks/kg_copilot.ipynb. In short: documents are read and cleaned,
entities are tagged and merged across languages, relationships are extracted by
a language model and then verified by an independent model before they enter
the graph, and the project's structured datasets are merged in as a factual
backbone. The full corpus run cost about $5 in model calls.