ksk-1729/freebase-neo4j-graph

Freebase → Neo4j Graph

A property-graph conversion of the final Freebase RDF dump (English-filtered), including proper resolution of Freebase's Compound Value Type (CVT) nodes, ready for import into Neo4j or use as a general-purpose large knowledge graph.

Freebase was a large collaborative knowledge base, discontinued by Google in 2016. This dataset is derived from the last publicly available RDF dump (freebase-rdf-latest.gz, 1.9B raw triples), filtered to English-language content (867M triples) and restructured from RDF into a clean property-graph shape.

Why this exists

Freebase's native RDF representation relies heavily on Compound Value Type (CVT) nodes — intermediate nodes used to represent n-ary facts (e.g. a film role links a film, an actor, and a character together through one CVT node). Naively importing Freebase's RDF into a property graph either leaves these as a confusing layer of indirection, or silently discards the information they carry. This dataset resolves that ambiguity explicitly, on a per-instance basis:

• 2-entity CVTs are collapsed into a single direct relationship between the two entities, with the CVT's other fields (dates, notes, etc.) becoming edge properties.
• 3+-entity CVTs are kept as their own nodes (true n-ary relations can't be cleanly collapsed to a single edge).
• CVTs with only one real entity connection have their fields attached directly as extra properties on that one entity, rather than being discarded.
• CVTs with no real entity connection are dropped.

Full methodology, including every pitfall hit along the way, is documented in CONVERSION_README.md in this repo.

Files

File	Rows (approx.)	Description
entity_nodes.parquet	~12.1M	Real-world entities (people, films, places, etc.)
cvt_nodes.parquet	~29.0M	Kept-as-node CVTs (3+-way compound facts)
entity_rels.parquet	~26.6M	Direct entity-to-entity relationships
cvt_direct_rels.parquet	~33.4M	Collapsed 2-way CVT relationships (with properties)
cvt_keepnode_rels.parquet	~174.7M	Edges connecting kept-CVT nodes to entities

Imported into Neo4j, this totals: 41,062,540 nodes, 187,981,682 relationships, 103,368,288 properties.

Schema

entity_nodes.parquet

Column	Description
mid:ID	Freebase machine ID (unique node identifier)
name	Primary display name (from type.object.name)
alias	Semicolon-separated list of known aliases
label:LABEL	Semicolon-separated Freebase type paths (e.g. people/person;film/actor)
attached_props	Semicolon-separated key=value pairs merged in from single-connection CVTs

cvt_nodes.parquet

Column	Description
mid:ID	Freebase machine ID
label:LABEL	Always CVT
props	Semicolon-separated key=value literal properties of the compound fact

Relationship files (entity_rels, cvt_direct_rels, cvt_keepnode_rels)

Column	Description
:START_ID	Source node mid
:END_ID	Target node mid
:TYPE	Relationship type, derived from the Freebase predicate (e.g. FILM_PERFORMANCE)
props (direct_rels only)	Semicolon-separated key=value edge properties

Reimporting into Neo4j

neo4j-admin database import full \
  --nodes=Entity="entity_nodes.csv" \
  --nodes=CVT="cvt_nodes.csv" \
  --relationships="entity_rels.csv" \
  --relationships="cvt_direct_rels.csv" \
  --relationships="cvt_keepnode_rels.csv" \
  --overwrite-destination \
  --skip-bad-relationships=true \
  --skip-duplicate-nodes=true \
  freebase-db

(Convert the Parquet files back to CSV first if needed — e.g. with DuckDB's COPY ... TO 'file.csv' (FORMAT CSV) — or adapt the import to read Parquet directly via a loader of your choice.)

Known limitations

• Freebase's type.object.key predicate was lost during English-language filtering (tagged via an older /lang/en convention not recognized by the filtering script used) and is not present anywhere in this dataset.
• A meaningful fraction of entities have no name property — this reflects genuine sparsity/incompleteness in the source Freebase data (many auto-generated or bulk-imported stub records were never given a curated name), not a conversion error.
• Some relationships in cvt_keepnode_rels.parquet reference mids that were intentionally excluded elsewhere in the pipeline (schema/meta objects); these are flagged for skipping during Neo4j import via --skip-bad-relationships.

License / attribution

Google distributed the Freebase Data Dumps under a Creative Commons Attribution (CC-BY) license, per the official Freebase data page ("Freebase Data Dumps are provided free of charge for any purpose with regular updates by Google. They are distributed, like Freebase itself, under the Creative Commons Attribution (aka CC-BY) license"). That page does not pin the data itself to a specific CC-BY version number, so this card follows that same generic "CC-BY" designation rather than asserting a version that isn't officially confirmed. (Note: the separate Freebase↔Wikidata ID mapping file, not included in this dataset, was released under CC0 instead.)

This derived dataset is distributed under the same CC-BY terms. If you reuse this data, please retain attribution to the original Freebase project, e.g.:

Derived from Google's Freebase Data Dumps (https://developers.google.com/freebase/), licensed CC-BY.

Citation

If you use this dataset, please cite the original Freebase project:

@misc{freebase:datadumps,
  title = "Freebase Data Dumps",
  author = "Google",
  howpublished = "\url{https://developers.google.com/freebase/data}",
}

and note that this repository is an independent, community-derived property-graph conversion, not an official Google/Freebase release.