jimfhahn/lc-bibframe-hubs

Library of Congress BIBFRAME Hubs — Parquet edition

A columnar, query-ready conversion of the Library of Congress BIBFRAME Hubs bulk export (hubs.bibframe.ttl.gz, 2026-05-05 publish). A Hub is the BIBFRAME equivalent of a MARC Title or Name/Title authority heading — an authorized access point such as Austen, Jane. Pride and prejudice or Homer. Odyssey. English given an HTTP URI so it can be linked to rather than string-matched. Hubs are lightweight aggregation nodes for collocating related resources; a single literary work may have several (the base work, a language expression, a spoken-word version…), and Hubs carry typed relationships to one another (translations, adaptations, sequels, parodies, film versions…). See Appendix A: More about Hubs in the BIBFRAME Primer. For how Hubs actually come into existence from MARC (240+100 AAPs, bare 130s, 758 Work links, series 7XX/8XX fields, and Modern MARC 240/130 $1 URIs), see the plugin's Modern MARC Hub Discovery notes — that taxonomy explains why a single title shows up here as several Hubs and why some marc_key values are name/title strings while others are bare titles.

The point of this dataset is to make the graph usable without a graph database: the whole thing is ~325 MB of Parquet that DuckDB (CLI, Python, or WASM in a browser) can query directly over HTTP range requests.

Hubs collocate records across classification

See it: Hubs collocate records across classification, an in-browser page built on the table below.

bib_to_hub.parquet links 493K Library of Congress bibliographic records to the Hubs here (437,836 at high confidence, 96.5% exactly right in a 300-record hand check). It shows what Hubs add to a catalog: classification puts a record on one shelf, while a Hub and its typed relationships pull together the novel, the opera, the score and the case study wherever each one is classed.

Record shelved in…	Hub relationship	…is now one hop from
Sacks, The man who mistook his wife for a hat — 6 records in R (medicine)	opera adaptation	Nyman's chamber opera — 2 records in M (music)
Chernow, Alexander Hamilton — E (American history)	inspiration for	Miranda, Hamilton — 2 records in M
Landon, Anna and the King of Siam — 4 records in D (history)	musical theatre adaptation	Rodgers, The King and I — 7 records in M
Cellini, Vita — 4 records in N (art)	opera adaptation	Berlioz, Benvenuto Cellini — 4 records in M
Herriman, Krazy Kat — 2 records in N	based on	Carpenter's Krazy Kat ballet — M

Classification puts each pair on different shelves. Subject headings rarely rejoin them: across all 40 cross-class pairs, the two sides share an identical LCSH heading in 3 and a main heading (usually a person or character, e.g. Hamilton, Alexander, Cellini, Benvenuto) in 14. Some records already name the related work in a name/title added entry (28 of 40 pairs), but only one-way and only on that record. The Hub graph links the works themselves, and bib_to_hub ties both ends to real records (https://id.loc.gov/resources/works/{lc_001}). At high confidence, 304K Hubs gather records, 69K of them gather two or more (up to 176 on one Hub), 2,303 gather records from two or more LC classes, and 298 creative relationships (adaptation, libretto, musical setting, based on…) have records on both ends, 40 of them across classes.

The records come from LC's experimental bibliographic embeddings, repackaged as jimfhahn/lc-bib-embeddings (12.2M records); lc_001 joins the two. That source has no class P (language and literature) or T yet (LC's files are not downloadable), and P is where most adaptations start, so these cross-class counts are a floor. Method, precision and caveats are in bib_to_hub.parquet details below.

-- every LC record on Sacks's Hub and on the Hubs one relation away
WITH me AS (SELECT 'a9402977-4b49-89c3-e703-6f4e0c969790' AS id),  -- Sacks, Man who mistook his wife for a hat
fam AS (
  SELECT id AS hub_id, 'this Hub' AS rel FROM me
  UNION
  SELECT CASE WHEN r.source_id = me.id THEN r.target_id ELSE r.source_id END, r.rel_type
  FROM 'hf://datasets/jimfhahn/lc-bibframe-hubs/relations.parquet' r, me
  WHERE me.id IN (r.source_id, r.target_id))
SELECT fam.rel, b.lcc_class, b.title, b.creator, b.lc_001
FROM fam JOIN 'hf://datasets/jimfhahn/lc-bibframe-hubs/bib_to_hub.parquet' b USING (hub_id)
WHERE b.confidence = 'high'
ORDER BY b.lcc_class;

Files

File	Rows	Size	One row per
hubs.parquet	2,907,948	227 MB	Hub (all of them), sorted by hub_id
hubs_connected.parquet	917,826	75 MB	Hub that appears in at least one relation — the browsable graph
relations.parquet	544,005	22 MB	Hub → Hub relation, sorted by (source_id, target_id)
rel_type_freq.parquet	146	tiny	relationship type, with count
bib_to_hub.parquet	493,331	35 MB	LC bibliographic record matched to a Hub (see details)
bib_to_hub_eval.csv	300	90 KB	hand-checked sample behind the precision figures
collocation/colloc_*.parquet	4 files	0.5 MB	small summaries (high confidence) behind the collocation page: records per Hub per class, creative Hub pairs, the records themselves, coverage by class

hubs*.parquet columns

column	type	notes
hub_id	string	LC UUID. Full URI is http://id.loc.gov/resources/hubs/{hub_id}
title	string	bf:mainTitle of the bf:Title (not variant) node; 100% populated
variant_titles	list<string>	bf:mainTitle of any bf:VariantTitle nodes (~1.05M Hubs have some)
marc_key	string	bflc:marcKey — the MARC 1XX/130/240-style authorized access point, e.g. 1001 $aAusten, Jane,$d1775-1817.$tPride and prejudice
lccn	string	bf:Lccn value if present (~1.58M Hubs)
agents	list<string>	contributor RWO URIs (http://id.loc.gov/rwo/agents/{naf_id}); ~1.89M Hubs have ≥1
media	list<string>	rdf:type local names other than Hub/Work: Series, Audio, NotatedMusic, MovingImage, Arrangement, NotatedMovement, NonMusicAudio, Multimedia, Text, MusicAudio

relations.parquet columns

column	type	notes
source_id	string	Hub that carries the bf:relation
target_id	string	Hub referenced by bf:associatedResource. 11,833 targets are not present in hubs.parquet (dangling in the LC dump)
rel_type	string	lower-cased local name of the relationship URI, e.g. translationof, adaptedasmotionpicture, parodyof. related when untyped
rel_type_uri	string	the raw relationship URI. Two vocabularies appear: http://id.loc.gov/vocabulary/relationship/… (most) and http://id.loc.gov/entities/relationships/…
typed	bool	false only when the source had no IRI-valued bf:relationship

Relations are recorded only on the source side in the LC data. To get a Hub's full neighborhood you must look in both directions (source_id = X OR target_id = X).

bib_to_hub.parquet details

One row per LC bibliographic record that matched a Hub, out of the 12.2M in lc-bib-embeddings.

column	type	notes
lc_001	string	MARC 001 of the LC record; its BIBFRAME Work is https://id.loc.gov/resources/works/{lc_001}; joins to lc-bib-embeddings
lccn, lcc_class, title, creator	string	copied from the record so the table is readable on its own
hub_id	string	best-matching Hub
match_method	string	aap, variant_title or uniform_title_130 (below)
confidence	string	high, medium or low (below)
name_dominated	bool	the record title is just the creator's name ("Picasso" by Picasso), or, for variant titles, the name plus one word against a collective Hub
series_title	bool	the Hub looks like a series title that collected different books (see below)
hub_qualified	bool	the Hub's access point carries $l/$k/$n/$p… (an expression, part or selection)
n_candidates, candidates	int, list	all equally ranked Hubs, most-connected first; hub_id is candidates[1]

How records were matched. Offline, on the authorized access point. The record's first creator (dates and fuller forms dropped) and title (subtitle and leading article dropped) are normalized and compared with each Hub's marc_key:

1. aap: creator + title = Hub $a + $t.
2. variant_title: creator + title = Hub $a + one of its variant_titles. This catches translations: Studies on the development of consciousness → Chesni, Recherches sur le développement de la conscience.
3. uniform_title_130: title only, against 130 Hubs, for records with no creator and at least two title words.

Records keep the best tier; among ties, unqualified Hubs win, then the most-connected one. Two error patterns found while checking are labelled down rather than dropped:

• Name-only titles → low. Exhibition catalogues called Picasso or Marc Chagall all match a generic Picasso, Pablo. Picasso Hub; only 5 of 40 checked were the Hub LC's own Work links.
• Series titles → medium. Dropping the subtitle collapses Chicken soup for the soul : … or The worst-case scenario survival handbook : … volumes onto one Hub. Such Hubs show a pattern editions of one work do not: a different subtitle on most records and records in three or more LC classes. 13 Hubs (106 records); the Constitution, the Federalist and Leviathan are not flagged.

Precision (300 records, stratified; each checked against the LC Work's own Hub links where it has any, otherwise by hand):

confidence	records	exactly the right Hub	right work
high: one candidate, tiers 1–2	437,836	96.5% (95% CI 93–98%)	99%
medium: several Hubs tie, or series title	30,425	66% (52–78%)	78%
low: 130 title-only, or name-dominated	25,070	66% (52–78%)	72%

The medium and low samples were drawn before the two rules above; the 4,700 name-only matches that moved into low are weaker than that sample suggests (5 of 40 confirmed by LC's Work links).

Use high as-is. Treat medium as "right work, Hub uncertain": the tied Hubs usually differ only by a qualifier (a dated vs. undated law, volume (Yoma) vs. (Ḥagigah)), so show candidates. Use low for exploration. Every judgment is in bib_to_hub_eval.csv.

What it complements. LC's BIBFRAME Work for a record links Hubs only when the MARC has a uniform title (240/130) or series; in an 80-record spot check, 7 Works had a bf:expressionOf Hub link. Most correct matches here are records whose Work links no Hub at all. Where LC does link one, it is often the language-expression Hub (Carmen. Italian & French) while bib_to_hub gives the base work, or the same Hub.

Limits. 3.6% of records match at high confidence (12% in M, 1.5–5% elsewhere): a Hub exists only where LC made a name/title authority, and translated titles match only through variant titles. Built from the 2026-05-05 Hubs dump and the 2025 embeddings release with build_bib_to_hub.py (about 2 minutes; reads only the text columns of the embeddings).

Top relationship types

translationof              436,332
related                     36,011
relatedwork                 25,011
arrangementof               23,673
containerof                  3,124
containedin                  3,104
basedon                      2,095
continuedby                  1,625
continuationof               1,589
musicformotionpicture        1,445
motionpictureadaptationof      951
inseries                       848

The long tail (adaptedasmotionpicture, operaadaptationof, parodyof, inspirationfor, derivative, sequel, …) is where the interesting cross-medium and creative-transformation links live.

Querying

DuckDB, straight from the Hub (no download)

-- everything one hop from Pride and Prejudice
WITH me AS (SELECT '013dc21c-e732-e589-24fe-bcc876264d3a' AS id)
SELECT r.rel_type,
       CASE WHEN r.source_id = me.id THEN 'out' ELSE 'in' END AS dir,
       h.title, h.media
FROM 'hf://datasets/jimfhahn/lc-bibframe-hubs/relations.parquet' r, me
JOIN 'hf://datasets/jimfhahn/lc-bibframe-hubs/hubs_connected.parquet' h
  ON h.hub_id = CASE WHEN r.source_id = me.id THEN r.target_id ELSE r.source_id END
WHERE r.source_id = me.id OR r.target_id = me.id
ORDER BY r.rel_type;

-- find a Hub by title + author
SELECT hub_id, title, marc_key
FROM 'hf://datasets/jimfhahn/lc-bibframe-hubs/hubs.parquet'
WHERE title ILIKE 'pride and prejudice' AND marc_key ILIKE '%Austen%';

Python

import duckdb
con = duckdb.connect()
con.sql("SELECT rel_type, n FROM 'hf://datasets/jimfhahn/lc-bibframe-hubs/rel_type_freq.parquet' LIMIT 20").show()

Browser (DuckDB-WASM / Observable)

The files are served with HTTP range-request support, so DuckDBClient.of({rels: FileAttachment(...)})-style loading or a plain read_parquet('https://huggingface.co/datasets/jimfhahn/lc-bibframe-hubs/resolve/main/relations.parquet') works client-side. relations.parquet + hubs_connected.parquet (~97 MB) is enough for a full interactive neighborhood explorer.

Provenance & method

• Source: https://id.loc.gov/download/resources/hubs.bibframe.ttl.gz (published 2026-05-05; ~735 MB gzipped, ~5.5 GB Turtle, ~152.6M triples).
• Conversion: ttl_to_parquet.py from the VuFind BIBFRAME Hub plugin repo. The dump is organised as self-contained # BEGIN … # END blocks per Hub, so each block is parsed independently with pyoxigraph and fanned out across processes; the full run takes about a minute.
• Only the predicates listed above are extracted. Admin metadata, notes, subjects, genre/form, identifiers other than LCCN, and non-Hub bf:associatedResource targets are dropped.
• Counts cross-check against an n10s/Neo4j import of the same dump (2,907,948 bf:Hub nodes; 532,172 Hub→Hub paths = 544,005 relations minus 11,833 dangling targets).

Caveats

• Point-in-time snapshot. LC republishes periodically; Hub URIs can drift. Verify a URI is live with a HEAD on https://id.loc.gov/resources/hubs/{hub_id}.rdf before linking.
• agents are bare RWO URIs — no names. Resolve via https://id.loc.gov/rwo/agents/{id}.json or the LC Names bulk files.
• Language tags on titles are not preserved.

License

The underlying data is published by the Library of Congress as a U.S. Government work and is in the public domain. This conversion is released under CC0 1.0.