sister-software/mailwoman-wof-gazetteer

Mailwoman WOF Gazetteer

Multi-locale administrative-hierarchy gazetteer for address parsing, derived from Who's On First. Ships a unified SQLite database plus seven per-locale FST (finite-state transducer) binaries for fast in-process gazetteer lookup.

• Source code: https://github.com/sister-software/mailwoman
• Build pipeline: WAL + Freeze design brief
• License: CC-BY-4.0 (inherited from upstream WOF)

Contents

admin-global-priority.db (1.09 GB)

Unified SQLite database for 7 priority countries. Schema mirrors the official geocode.earth WOF distribution for drop-in compatibility.

Table	Rows
spr (single-place records)	1,288,749
names (multi-language variants)	10,233,886
concordances (cross-source IDs)	2,156,440
place_population	220,469

Countries: US (449K places), CN (680K), FR (231K), DE (189K), GB (73K), JP (63K), KR (54K).

fst/fst-{locale}.bin — per-locale FST gazetteers

Pre-built FST gazetteers for use as Viterbi emission priors in the mailwoman neural classifier. Each FST is built from the global SQLite with all-language name variants (CJK, Cyrillic, Arabic, Thai work).

Locale	File	Size	States	Places
en-US	fst-en-us.bin	20.9 MB	160K	449K
en-GB	fst-en-gb.bin	3.7 MB	33K	73K
fr-FR	fst-fr-fr.bin	10.2 MB	72K	231K
ja-JP	fst-ja-jp.bin	13.0 MB	116K	63K
ko-KR	fst-ko-kr.bin	7.1 MB	55K	54K
zh-CN	fst-zh-cn.bin	92.5 MB	589K	680K
de-DE	fst-de-de.bin	8.1 MB	70K	189K

Binary format: V4 (u32 edge/place counts per state). Decoder: @mailwoman/resolver-wof-sqlite/fst-serialize.

Multi-script support

Token normalization uses Unicode property escapes (/[\p{P}\p{S}]/gu) instead of ASCII-only patterns. CJK queries work:

Query	Top match
東京	region Tokyo (importance 0.98)
北京	region Beijing (1.00)
서울	region Seoul (0.95)
大阪	region Osaka (0.94)
パリ	region Paris (0.79)
москва	locality Moscow

Build provenance

Field	Value
Source	Who's On First admin repos (cloned via gh repo list whosonfirst-data --limit 1000)
Build script	scripts/build-unified-wof.ts
Build pipeline	WAL → checkpoint → DELETE → ANALYZE → VACUUM INTO
GeoJSON files processed	1,743,963 (alt-geometries excluded)
Build time	185s on a single workstation

Usage (JavaScript)

import { deserializeFst } from "@mailwoman/resolver-wof-sqlite/fst-serialize"
import { readFileSync } from "node:fs"

const buf = readFileSync("./fst-en-us.bin")
const matcher = deserializeFst(buf)
const result = matcher.query("new york")
// → { accepting: [{ placetype: "region", name: "New York", importance: 1.0 }, ...] }

Usage (SQLite direct)

SELECT s.name, s.placetype, s.latitude, s.longitude
FROM spr s
JOIN names n ON n.id = s.id
WHERE n.language = 'jpn' AND n.name = '東京'
LIMIT 5;

Updates

Rebuilt periodically as WOF upstream changes. Tagged versions track the source-repo commit hashes at build time.

License

CC-BY-4.0. When redistributing, attribute Who's On First and link to https://whosonfirst.org/.