Five marisa-trie native binaries that support multilingual entity linking with mGENRE. They are the runtime assets for the impresso-project/nel-mgenre-multilingual model as run by the impresso-inference harness:

a title prefix tree that constrains beam search to valid Wikipedia titles,
a (language, title) → Wikidata QID lookup for offline QID resolution — GENRE's original, and a copy with every QID merged on Wikidata since replaced by its current target (the one used at inference),
a QID → (facets, birthdate) attribute table for offline post-linking filters (disambiguation / name-page drop, birthdate anachronism, type mismatch, entity-type repair), and
a QID → labels table (French / English / German Wikidata labels) to enrich linked entities with human-readable names.

All are mmap-loaded at startup, so cold start is seconds regardless of file size (the legacy pickle path took ~57 min).

Files

File	Size	Format	Role
`titles_lang_all105_marisa_trie_with_redirect.marisa`	~600 MB	`marisa_trie.Trie` (native `.save`/`.mmap`)	Prefix tree of valid `Title >> lang` token-id sequences (105 languages, redirects included). Drives constrained beam search via a `prefix_allowed_tokens_fn` callback.
`lang_title2wikidataID-normalized_with_redirect.marisa`	~850 MB	`marisa_trie.BytesTrie`	Raw conversion of GENRE's lookup: `f"{lang}\x1f{title}"` → the lex-smallest Wikidata QID (ASCII), as of the ~2020 GENRE snapshot. Input of the resolved lookup below; not read at inference.
`lang_title2wikidataID.qid-resolved.marisa`	~850 MB	`marisa_trie.BytesTrie`	Same keys; every QID merged on Wikidata since the GENRE snapshot replaced by its current target. Resolves generated titles to QIDs offline.
`qid2attrs.v2.marisa`	~80 MB	`marisa_trie.RecordTrie` (fmt `"<HI"`)	Maps each Wikidata `QID` → `(facet_flags, birth_days)`. Per-QID attributes for offline post-linking filters. Supersedes `qid2attrs.marisa` (v1, `"<BI"` single class code), kept for older releases.
`qid2labels.marisa`	a few hundred MB	`marisa_trie.BytesTrie`	Maps each Wikidata `QID` → its packed `fr` / `en` / `de` labels (+ one fallback label). Per-QID names for output enrichment.

The three GENRE-derived files hold ~89 million keys each (all105 = the 105 languages of the underlying mBART-50 / mGENRE vocabulary); qid2attrs.v2.marisa holds ~20.2 million keys — one per distinct QID reachable through the lookup — and qid2labels.marisa one per such QID that still has a label on Wikidata.

Provenance & construction

These files are converted, byte-faithfully, from Facebook Research's GENRE public downloads at https://dl.fbaipublicfiles.com/GENRE/:

titles_lang_all105_marisa_trie_with_redirect.pkl (~582 MB) — a pickled GENRE MarisaTrie wrapper. Conversion extracts the inner marisa_trie.Trie and re-saves it as a bare native binary so it can be mmap-ed directly.
lang_title2wikidataID-normalized_with_redirect.pkl (~3.9 GB) — a dict[(lang, title), str | set[QID]]. Conversion streams it into a marisa_trie.BytesTrie, collapsing multi-QID values to the lex-smallest QID at build time (so the runtime does a single trie lookup instead of an ~8-minute normalisation pass per launch).

The conversion is performed by the impresso-nel-stage-assets tool in impresso-inference (src/impresso_inference/tasks/nel/cli/stage_assets.py). Nothing about the entity inventory is added or removed — this is a format/packaging re-host of the GENRE data for fast startup.

Resolved lookup (`lang_title2wikidataID.qid-resolved.marisa`)

About 2.8 % of the GENRE lookup's distinct QIDs (1.2 % of its keys) have since been merged into another Wikidata item; the old id survives only as an owl:sameAs redirect, with no class, birthdate or label. The impresso-nel-stage-lookup tool (src/impresso_inference/tasks/nel/cli/stage_lookup.py) asks the QLever Wikidata endpoint for those redirects (batched ?item owl:sameAs ?target), follows chains (up to 5 hops; a cycle leaves the QID as is), and rewrites each merged value to its current target. Keys are unchanged; QIDs deleted outright are kept as they are. The build is rejected if the key count changes, if a sampled value is still a redirect, or if more than 10 % of keys would be rewritten. The attribute and label tables below are built from this file, so they are keyed by current QIDs only.

QID attributes (`qid2attrs.v2.marisa`)

Unlike the two files above, qid2attrs.v2.marisa is not derived from GENRE — it is built from Wikidata so that post-linking filters run fully offline (no live Wikidata/Wikipedia HTTP calls). It is keyed by Wikidata QID and each value is a fixed 6-byte record (marisa_trie.RecordTrie, struct format "<HI"), plus one non-QID row __format__ → (2, 0) identifying the layout:

Field	Type	Meaning
`facet_flags`	`uint16` (`H`)	Bitmask of every facet the QID reaches (a QID can have several): bit 0 `pers`, 1 `org`, 2 `loc`, 3 `disambig`, 4 `name` (given/family name), 5 `periodical`, 6 `doctr` (religious text / law / treaty), 7 `time_cal` (year, century, decade, millennium, calendar day, weekday, holiday), 8 `period_event` (time interval / occurrence), 9 `vehicle`, 10 `taxon`. Bits 11–15 unused.
`birth_days`	`uint32` (`I`)	Earliest date of birth as days since `0001-01-01` (proleptic Gregorian); always `>= 0`. Sentinel `2147483648` (`2*31`) means no/unusable birthdate* (absent, BCE, or imprecise).

Construction — by the impresso-nel-stage-qid-attrs tool (src/impresso_inference/tasks/nel/cli/stage_qid_attrs.py):

Enumerate the linkable universe = the distinct QID values of lang_title2wikidataID.qid-resolved.marisa (the only QIDs mGENRE can emit).
Pull attributes from the QLever Wikidata endpoint, in batches of 10,000 QIDs (VALUES ?item; each query stays under the public instance's 30 s limit, and a truncated result aborts the build):
- facets via wdt:P31/wdt:P279* (or direct wdt:P31) against fixed roots — person wd:Q5; organization wd:Q43229; location wd:Q27096213 + wd:Q42124; disambiguation pages via direct wdt:P31 on the six page types (incl. human-name wd:Q22808320 and surname wd:Q66480449); name wd:Q202444 + wd:Q101352; periodical wd:Q1002697; doctrine wd:Q179461 + wd:Q7748 + wd:Q131569; calendar wd:Q577 + wd:Q578 + wd:Q39911 + wd:Q36507 + wd:Q14795564 (+ direct wd:Q41825); period/event wd:Q186081 + wd:Q1190554; vehicle wd:Q42889; taxon direct wd:Q16521;
- birthdate via the earliest wdt:P569 (conservative: only the earliest claim).
Pack one record per linkable QID.

Role. In impresso-inference the NEL writer looks up each linked QID here to drop links that cannot be right — disambiguation pages, a person linked to a name item, a person born after the article's publication date, and links whose QID is incompatible with the mention's entity type (a date linked to a place, a newspaper title linked to its city) — and to repair pers/org/loc entity types (e.g. an org-tagged mention whose QID is a location).

QID labels (`qid2labels.marisa`)

Also built from Wikidata (via QLever), keyed by the same linkable QIDs. Each value is the UTF-8 string of 5 fields separated by \x1f (an empty field = absent):

#	Field	Meaning
1	`label_fr`	French label, else the item's `mul` label, else empty
2	`label_en`	English label, else `mul`, else empty
3	`label_de`	German label, else `mul`, else empty
4	`label_other`	Set only when fields 1–3 are all empty: the label in the item's own language
5	`label_other_lg`	Language code of `label_other` (e.g. `ja`, `lb`, `ru`)

Fallback rules. mul is Wikidata's "default for all languages" label (introduced in 2024; many items now carry a name only there). A language field never takes another language's label — only its own or mul. label_other is the label in a language whose Wikipedia title maps to the QID in the lookup above (smallest code if several), else the smallest language code with a label. QIDs with no label at all — in practice items deleted on Wikidata since the GENRE snapshot (merged ones are already resolved) — are not stored; a lookup miss means "no label".

Coverage (estimate from a 20,000-QID sample of the linkable set, batched build of 2026-10-02; refreshed on the full build): label_fr 74 %, label_en 91 %, label_de 70 % (each incl. ~1.5–3 % filled from mul); label_other on ~8 %; ~1.5 % with no label.

Construction — by the impresso-nel-stage-qid-labels tool (src/impresso_inference/tasks/nel/cli/stage_qid_labels.py): rdfs:label queries in batches of 10,000 linkable QIDs (VALUES ?item, languages fr, en, de, mul); then batched queries for all labels of the QIDs still empty, to pick label_other. Each query stays under QLever's 30 s limit, a truncated result aborts the build, and the build is rejected if fr/en/de coverage falls below 40 / 70 / 35 %.

Note (2026-10-02). The first qid2labels.marisa published here (2026-10-01) was truncated (en ~15 % coverage) by QLever's 30 s limit, and is replaced by the batched rebuild.

Wikidata snapshot. The resolved lookup, qid2attrs.v2.marisa and qid2labels.marisa reflect the QLever Wikidata index at the time they were built — see each file's commit date in this repo.

Rebuild

Each file has exactly one build step, one console script, and one make target in impresso-inference:

Step	Console script	`make` target (CPU-only RCP pod)	Builds	Needs
1	`impresso-nel-stage-assets`	`runai-stage-nel-assets`	trie + GENRE lookup	GENRE downloads
1b	`impresso-nel-stage-lookup`	`runai-stage-nel-lookup`	resolved lookup	step-1 lookup + QLever
2a	`impresso-nel-stage-qid-attrs`	`runai-stage-nel-qid-attrs`	`qid2attrs.v2.marisa`	step-1b lookup + QLever
2b	`impresso-nel-stage-qid-labels`	`runai-stage-nel-qid-labels`	`qid2labels.marisa`	step-1b lookup + QLever
all	—	`runai-stage-nel-all`	everything, 1 → 1b → 2b → 2a	—

Steps 2a and 2b are independent of each other; rebuild them whenever 1b is rebuilt. Every tool skips an existing output (--force to rebuild), supports a quick --limit N dry run, and uploads with --push.

Intended use

Pair with impresso-project/nel-mgenre-multilingual (mGENRE, an mBART-50 sequence-to-sequence entity linker). mGENRE emits strings of the form "Wikipedia_Title >> xx":

The trie restricts generation to valid titles at every decoder step (prevents hallucinated pages).
The lookup turns the generated (title, lang) into a Wikidata QID without any live Wikipedia/Wikidata HTTP calls — suitable for offline / batch inference.
The attributes drive offline post-linking filters; the labels give each linked entity its French / English / German name (label_fr / label_en / label_de).

In impresso-inference, the NEL task auto-downloads whichever runtime file is missing from this dataset on first launch (qid2labels.marisa is staged ahead of its runtime use in the NEL writer). See that repo's src/impresso_inference/tasks/nel/ for the full pipeline.

How to load

import marisa_trie

# Title trie — constrained decoding
trie = marisa_trie.Trie()
trie.mmap("titles_lang_all105_marisa_trie_with_redirect.marisa")

# (lang, title) -> QID lookup
lookup = marisa_trie.BytesTrie()
lookup.mmap("lang_title2wikidataID.qid-resolved.marisa")
qid = lookup[f"en\x1fGermany"][0].decode("ascii")   # -> "Q183"

# QID -> (facets, birthdate) attributes
from datetime import date, timedelta
FACETS = ["pers", "org", "loc", "disambig", "name", "periodical", "doctr",
          "time_cal", "period_event", "vehicle", "taxon"]   # bit i = FACETS[i]
BIRTH_NONE, EPOCH = 2**31, date(1, 1, 1)
attrs = marisa_trie.RecordTrie("<HI")
attrs.mmap("qid2attrs.v2.marisa")
flags, birth_days = attrs["Q42"][0]                 # Douglas Adams
facets = [f for i, f in enumerate(FACETS) if flags >> i & 1]   # -> ["pers"]
birth = None if birth_days == BIRTH_NONE else EPOCH + timedelta(days=birth_days)  # -> 1952-03-11

# QID -> labels (5 fields, "" = absent)
labels = marisa_trie.BytesTrie()
labels.mmap("qid2labels.marisa")
hit = labels.get("Q183")                             # None if the QID has no label
fr, en, de, other, other_lg = hit[0].decode("utf-8").split("\x1f")
# -> "Allemagne", "Germany", "Deutschland", "", ""

Token-id encoding (trie only). marisa-trie stores keys as UTF-8 strings, so mGENRE token ids are encoded per character as chr(t) for t < 55000 and chr(t + 10000) otherwise — a 10 000-codepoint shift that skips the UTF-16 surrogate range. This matches the upstream GENRE byte layout; any consumer of the trie must apply the same shift when decoding allowed-token sets. Keys are Title >> lang token sequences.

Licensing

Released under CC BY-NC 4.0 (non-commercial), inherited from the upstream facebookresearch/GENRE data the trie and lookup are derived from. The underlying knowledge base is public: Wikidata identifiers are CC0, and Wikipedia titles are CC BY-SA. Use for non-commercial research consistent with the GENRE license.

qid2attrs.v2.marisa and qid2labels.marisa contain only Wikidata data (P31/P279/ P569, rdfs:label), which is CC0; their QID selection follows the GENRE-derived lookup above, so they are distributed here under the same non-commercial terms for consistency.

Citation

The trie and lookup originate from mGENRE:

@article{de-cao-etal-2022-multilingual,
  title   = "Multilingual Autoregressive Entity Linking",
  author  = "De Cao, Nicola and Wu, Ledell and Popat, Kashyap and Artetxe, Mikel
             and Goyal, Naman and Plekhanov, Mikhail and Zettlemoyer, Luke and
             Cancedda, Nicola and Riedel, Sebastian and Petroni, Fabio",
  journal = "Transactions of the Association for Computational Linguistics",
  volume  = "10",
  year    = "2022",
  address = "Cambridge, MA",
  publisher = "MIT Press",
  url     = "https://aclanthology.org/2022.tacl-1.16",
  doi     = "10.1162/tacl_a_00460",
  pages   = "274--290"
}

Acknowledgements

Repackaged for historical-newspaper entity linking by the Impresso project — an interdisciplinary effort on historical media analysis across languages, time, and modalities. Funded by the Swiss National Science Foundation (CRSII5_173719, CRSII5_213585) and the Luxembourg National Research Fund (grant No. 17498891).