Miñano · Balearic subset

Live at https://minano-balears.cloudflare-d82.workers.dev/.

Digital edition of the Balearic Islands articles of Sebastián Miñano y Bedoya's Diccionario geográfico-estadístico de España y Portugal (Madrid, 1826–1829, 11 vols.).

The original work is the first modern geographical-statistical dictionary of the Iberian Peninsula and the immediate predecessor of Pascual Madoz's Diccionario (1845–1850). It describes a pre-liberal Spain that still records the seigneurial regime of every locality (realengo, señorial, abadengo, de órdenes, eclesiástico), two decades before Mendizábal's confiscations.

This repository extracts every article relating to Mallorca, Menorca, Eivissa, Formentera and Cabrera, structures the data into a relational schema, and publishes a static website for consultation and exploration.

Coverage

Indicator	Value
Volumes processed	11 / 11 (1826–1829)
Balearic articles extracted	182
Articles with geographic coordinates	99.5 %
Public website	web/ (static, vanilla JS)
License	AGPL-3.0-or-later (code); original text in the public domain

The only documented gap is Tomo X (VIL–VIZ): the available Internet Archive scan is partial and does not include any Balearic entry.

Pipeline

The extraction is a four-stage pipeline. All stages are deterministic except the Anthropic-model call in stage 2; the model output is checked into the repository as JSON, so the database and the website can be rebuilt without re-spending tokens.

1. OCR ingestion. For each volume, the chOCR (character-level hOCR with bounding boxes) is fetched from Internet Archive. Tomos V and X are not served as chOCR by Internet Archive; they are regenerated locally from the ABBYY FineReader XML using the same open-source toolchain that Internet Archive itself runs. The result is bit-for-bit equivalent to what IA serves for the other volumes.

2. Article identification and extraction. A fuzzy regex indexer scans the chOCR paragraph by paragraph and selects those that contain a Balearic anchor (place name plus jurisdictional marker). A second, OCR-tolerant pass recovers articles that the first regex missed because of severe OCR damage (e.g. POLUSNZA for Pollença, ALClülA for Alcúdia, CABRER^A for Cabrera). Each candidate leaf is then sent to an Anthropic language model (current run: Claude Opus 4.7), which reads the chOCR text and emits a structured JSON object per article with the following fields:

   title · place_type · island · seigneurial_regime · mayor_type
   municipality · description · stats · cross_references · confidence
   

   The extracted JSONs live under data/text/page_<vol>_<leaf>.json and are the canonical source of truth for the rest of the pipeline.

   The municipality field follows a consistent convention: for villes and ciutats — which are themselves their own municipality — it carries the modern Catalan form of the title (e.g. ALAYOR → Alaior, MAHON → Maó, PUEBLA (la) → sa Pobla); for subordinate entries (lugar, aldea, caserío, cortijo, santuari, castell…) it carries the historical parent municipality at the time Miñano was writing (so for instance Llorito points to Sineu, of which it was then an annex, rather than to the modern stand-alone municipi of Lloret de Vistalegre). Supramunicipal entities — islands, termes menorquins, quartons eivissencs and natural features (caps, ports, serres, valls) — leave the field empty, which is the semantically correct reading of "this place has no single parent municipality".

   A complementary attribution-based recovery pass catches articles whose title was so OCR-damaged that no place-name regex could find them. Fra Lluís de Vilafranca, the Capuchin friar of Palma who served as Miñano's Balearic correspondent, signed the great majority of Balearic adicions in the Suplemento with the formula «Not. dada por el R. P. Fr. Luis de Villafranca». The script scripts/vilafranca_recovery.py scans every chOCR volume looking for paragraphs whose tail fuzzy-matches that attribution, filters out the many peninsular false positives produced by the Catalan corregimiento de Villafranca district, and reports paragraphs absent from the corpus. The signature alone is not sufficient (a peninsular paragraph may quote Vilafranca in passing), so a Balearic-keyword guard rules out the corregimiento mentions; in exchange, the matcher does not depend on the lemma being readable at all, which is what allowed the recovery of entries like Ullaró (printed by Miñano under a heading rendered «LLtKO. AH. H.» in the raw OCR). This pass added fifteen Suplemento adicions that the place-name indexer had missed.

   A second recovery layer — the buried-lemma audit (scripts/inner_lemma_audit.py) — looks for Balearic articles that the indexer overlooked because they were not at the start of their chOCR paragraph. Two distinct miss patterns are scanned. Pattern A (tail-merge) catches full article openers («TITLE, V.|L.|C.|Ald. de Esp. en la isla de Mallorca…») appearing mid-paragraph after the tail of an unrelated (usually peninsular) entry — the cases of SINEU (Tom VIII, merged with the tail of SINES of Portugal), RUBERTS (Tom VII, merged with QUINTANAPALLA), GALILEA (Tom XI, merged with the tail of a Galician parish) and GENOVA (Tom XI, merged with a Málaga-province editorial correction) were all recovered this way. Pattern B (chained Suplement corrections) catches short adicions of the form «Title. Tiene 733 vec…» chained one after another in a single Suplement paragraph; the indexer would typically capture only the first item in the chain and miss the rest, as happened with the Sencelles–Binissalem corrections paragraph (BINIARAUS, BINIALI, BINIAMAR, BINIARAIG) and the ALARÓ adición. The audit bounds each candidate's correction body to the next chained correction and requires a strict Balearic anchor (Mallorca, Menorca, Eivissa, Formentera, isla de Cabrera, las Baleares) or the Vilafranca attribution within that body, so peninsular adicions whose paragraph happens to share a Vilafranca signature with a Balearic neighbour do not contaminate the result.

   A third audit (scripts/suspicious_titles_audit.py) operates on the extracted titles themselves, looking for the OCR confusions documented in this corpus: BIM- or BUM- / BUNI- openings where Mallorcan toponymy expects BINI- (NI→M, NI→U); the letter K (almost always a misread R); the diaeresis on Ü outside the canonical Catalan üe/üi digraphs; and lemmas whose best fuzzy match in the gazetteer falls below a configurable WRatio threshold. Findings are split into probable issues (low fuzzy score, with or without a heuristic flag) and archaic-spelling notes (OCR-looking but matched ≥ 92 against NGIB — e.g. BÜÑOLA matching modern Buñola, an intentional 19th-century spelling rather than scanner damage), so a human pass can resolve the former without reviewing the latter.

3. Geocoding. Each extracted title is matched against the Nomenclàtor Geogràfic de les Illes Balears (NGIB, Govern de les Illes Balears: 55 531 modern toponyms). The geographic data download is fully reproducible from the IDEIB ArcGIS REST endpoint. The matcher normalises Miñano's hispanicised orthography (Iviza → Eivissa, Bañalbufar → Banyalbufar, hyphens and en/em dashes folded to spaces so ALCARIA-ROJA aligns with ALQUERIA ROJA) and curates a short table of 19th-century Castilian forms whose modern equivalents NGIB carries under a different lemma (e.g. Alcaria → Alqueria).

   Resolving toponyms against NGIB faithfully proved more involved than a single fuzzy lookup. The matcher therefore relies on six complementary safeguards, each addressing a distinct failure mode observed in NGIB's structure or in Miñano's typography:

   • Authoritative-type priority. Many Balearic toponyms recur in NGIB under several local_type values (e.g. sa Pobla exists both as a Municipi at the northern village and as a Finca, possessió at a Llucmajor farm). The gazetteer builder sorts candidates by an explicit priority list — Municipi > Capital de municipi > Vila > Entitat de Població > Llogaret > … > Finca, possessió — so the most authoritative row wins the (normalised, island) deduplication.

   • Complete settlement-type coverage. NGIB classifies many small but real settlements (Fornells of Menorca, es Capdellà, Llucalcari, several Eivissan parish villages) under the type Altre nucli de població, llogaret, and the four main islands and their islets under Illa gran and Illa mitjana. All these are now included in the gazetteer; their previous omission was causing matches to fall through to homonyms on other islands.

   • Strict same-island matching. When an entry declares its island, the candidate pool is restricted to that island. Cross-island fallback only applies when no island is declared (e.g. Balearic-wide articles). Without this constraint Eivissa's es Fornells was wrongly matched against Menorca's Fornells, Mallorcan possessions against Menorcan homonyms, and so on.

   • Length-disparity guard. Levenshtein-based fuzzy scorers (such as WRatio) award high marks when one string is a substring of the other; without correction, ROJA scored 90 against ALCARIA ROJA. Candidates shorter than 60 % of the title's normalised length are discarded before scoring.

   • Municipality tiebreaker. When the article declares a parent municipality (Sit. al N. O. de Palma…), homonyms in that municipality are preferred over equally-scoring candidates elsewhere on the same island.

   • Curated overrides and ambiguous-homonym signalling. A small table of explicit (island, title) → (lon, lat) entries resolves cases NGIB cannot disambiguate algorithmically because the title appears under several settlements with different types (e.g. la Vileta of Palma over the four other Mallorcan sa Vileta hamlets). A second table marks titles whose Miñano description is too thin to permit any disambiguation — the one-line Coto Redondo notices that name no village, no distance and no jurisdiction — and routes them to the island centroid with a distinct fallback label so they are visually distinguishable from centroid-fallback cases of unmatched toponyms.

   Articles that cannot be matched directly fall back to their parent municipality, and articles with no resolvable parent fall back to the centroid of the corresponding island. The aggregate match rate is 99.5 %; every matched entry is independently verified to land within the bounding box of its declared island and within 1 km of the canonical NGIB coordinates of its matched toponym.

4. Publication. The per-article JSONs are loaded into a DuckDB database and exported as a single web/data.json consumed by the static site. The site uses no framework: HTML, CSS and one JavaScript file. Leaflet is loaded lazily from CDN on the Mapa tab.

Website

The static site under web/ presents the corpus through five tabs:

• Inici — orientation, statistics by island, source-volume table.
• Explorar — filtered table of every article, with full text on expansion and a direct link to the corresponding page of the Internet Archive facsimile.
• Mapa — geographical distribution on a Leaflet map; circle size proportional to declared vecinos, colour by island.
• Estadístiques — nine visualisations covering demography (habitants and vecinos by entry, household size), fiscality (riqueza líquida in libras mallorquinas), administrative typology, per-volume distribution, and a sunburst diagram cross-tabulating island, seigneurial regime and place type.
• Notes — scholarly notes on Miñano's biography, the method by which the Diccionario was elaborated (correspondence with parish priests, after the model of the Felipe II Relaciones Topográficas), the role of fra Lluís de Vilafranca as the Balearic correspondent, the Corrección fraterna polemic with Fermín Caballero, the transition to Madoz, and the abbreviations used in the original work.

The website language is Catalan. Code, scripts and this README are in English.

Sources

The eleven volumes are scattered across multiple Internet Archive deposits — different libraries, no shared identifier prefix. The canonical source used for each volume is listed in the Fonts section of web/index.html (Tom I to XI, with the digitising institution and the IA identifier of each scan).

Running locally

uv venv && uv pip install -e .

# Reproduce the full extraction (writes data/text/, db/minano.duckdb,
# web/data.json). Requires ANTHROPIC_API_KEY for stage 2.
python scripts/fetch_volume.py <vol>          # OCR ingestion (one volume)
python scripts/derive_chocr_from_abbyy.py 05  # for Tomos V and X only
python scripts/index_volume.py <vol>
python scripts/extract_text.py --all <vol>    # stage 2 (LLM)
python scripts/fetch_ngib.py                  # NGIB gazetteer download
python scripts/build_gazetteer.py
python scripts/enrich_coords.py               # stage 3 (geocoding)
python scripts/load_text.py                   # stage 4 (DB)
python scripts/export_web_data.py             # stage 4 (web/data.json)

# Serve the static site
python -m http.server -d web 8000

The pipeline is idempotent: re-running any stage overwrites its outputs in place, with the data under data/text/ as the single source of truth.

License

Code: AGPL-3.0-or-later (see LICENSE). Running a modified version as a network service obliges the operator to make the modifications available to its users.

Original text (1826–1829) and the Internet Archive facsimiles derived from public-domain editions are themselves in the public domain.