TigreGotico/dicionario_barranquenho

Dicionário de Barranquenho

Structured lexical dataset derived from the first published dictionary of Barranquenho, a Romance contact language spoken in Barrancos, Portugal. Contains 1,680 entries with Portuguese and Spanish glosses, grammatical categories, semantic fields, source attributions, and synonym cross-references.

Language

Barranquenho (glottocode: barr1245; no ISO 639-3 code assigned at time of publication) is a contact language spoken in the municipality of Barrancos, Baixo Alentejo, Portugal (~1,500 inhabitants). It is of Portuguese base with systematic phonological, morphosyntactic, and lexical features from southern Spanish (Extremaduran and Andalusian varieties), along with archaisms and Leonese traces. The community is effectively trilingual in Barranquenho, Portuguese, and Spanish.

Barranquenho was legally recognised as cultural heritage by the Portuguese Parliament (Lei n.º 97/2021, 30 December 2021). The first orthographic convention, grammar, and dictionary were published in 2025.

Source

This dataset is derived from:

Gonçalves, Maria Filomena; Navas, María Victoria; Ferreira, Vera (2025). Dicionário de Barranquenho. 1ª ed. Évora: Universidade de Évora. ISBN 978-972-778-460-8

Authors' institutional affiliations:

• —Maria Filomena Gonçalves — Universidade de Évora / CIDEHUS-UÉ / FCT / Cátedra UNESCO em Património Imaterial e Saber-Fazer Tradicional
• —María Victoria Navas — Universidad Complutense de Madrid / CLUL
• —Vera Ferreira — CIDLeS – Centro de Documentação Linguística e Social

The dictionary was produced under a programme of preservation and valorisation of Barranquenho coordinated by the Câmara Municipal de Barrancos, funded through the FCT project UIDB/00057/2025.

This dataset is a derived work. Users are responsible for ensuring their use complies with the terms of the original publication and applicable copyright law. If you use this dataset in academic work, cite both the original dictionary and this dataset.

Dataset creation methodology

The source dictionary was published as a PDF with rasterised (image-only) pages. No machine-readable text was available. The following pipeline was used to produce this dataset:

1. Rasterisation

All 122 pages of the PDF were rendered to PNG images at 300 DPI using pdftoppm (poppler-utils). Dictionary entries occupy pages 28–122 (95 pages).

2. Column segmentation

Each dictionary page uses a two-column layout with a vertical rule at the page centre (~x = 877 px at 300 DPI). Each page was cropped into left and right halves (with a 15 px gap each side of the divider to exclude the rule), and each half was OCR'd independently.

3. OCR

Columns were processed with Tesseract 5 using the combined por+spa language model and --psm 6 (assume uniform block of text). Page 84 required special handling: the right column contained a full-page botanical photograph that was excluded by cropping only the top and bottom strips of that column.

4. Normalisation

OCR output was normalised to correct systematic recognition errors before parsing:

5. Entry parsing

Entry boundaries were detected with a regex anchored to line-start headwords followed by a grammatical category marker (n.f, n.m, v, adj, adv, expr, top, antr, etc.). Fields were extracted with per-field regexes:

• —pt. → pt_gloss (stops at esp., sin., {Uso, (Campo, [Fonte)
• —esp. → es_gloss (same stop conditions)
• —{Uso: ...} → uso
• —(Campo semântico: ...) → campo_semantico
• —[Fonte: ...] → fonte
• —sin. ... → sinonimos
• —{var. de ...} → variante_de
• —Multi-sense entries (marked 1. pt. ... 2. pt. ...) were split into separate rows

Entries were grouped by the first letter of the headword (Unicode-normalised) and deduplicated on (headword.lower(), category, sense), keeping the most field-complete instance where duplicates arose from two-column boundary overlap.

6. Validation and cleaning

A multi-pass validation was applied:

• —Column-count integrity (all rows must have exactly 11 fields)
• —Headword sanity (no empty, no embedded pt./esp. markers, no all-caps OCR blobs)
• —Gloss-bleed detection: es_gloss or pt_gloss containing a subsequent entry's category marker (n.f ... pt., loc.adv. pt., »n.f., etc.) — truncated at the bleed point
• —OCR residue detection (pí., vf., ad).) in gloss fields — re-extracted or manually corrected
• —One corrupted row removed (headword was esp. aceite. azeti de premêra — pure column-boundary noise)
• —ropéru sense 2 gloss recovered manually by cross-referencing the adjacent right-column OCR

The audio-marker column (indicating entries with an associated audio recording in the original digital resource) was removed as it was unreliable after OCR normalisation.

Schema

Grammatical categories

Statistics

Counts

Categories

Gloss coverage

The majority of entries have a single-word gloss in each language. Longer glosses occur for culturally-specific items with no direct equivalent (e.g. iguala — a traditional payment agreement between rural workers and doctors).

Semantic fields

Sources

Source codes: CLUL, I / II = Atlas Linguístico-Etnográfico de Portugal e da Galiza fieldwork data held at the Centre of Linguistics, University of Lisbon; Clements = Clements (2009), The Linguistic Legacy of Spanish and Portuguese; Valério = Valério (2023), Retalhos da vida do povo de Barrancos; ALEPG = Atlas Linguístico-Etnográfico de Portugal e da Galiza.

Orthographic features in headwords

These counts illustrate the systematic phonological differences between Barranquenho and standard Portuguese visible in the headword orthography.

Attested special characters in headwords: à á â ã ç é ê í ó ô ú

Data quality and applied corrections

This section documents every class of correction applied to the raw OCR output, distinguishing between legitimate OCR character fixes, structural cleaning, and residual issues that could not be resolved without access to the print source.

OCR character substitutions (applied, high confidence)

These are 1-to-1 character confusions produced by Tesseract on this particular typeface. All are confirmed by examining multiple instances.

Entry-boundary bleed truncation (applied)

The two-column page layout caused the right-column OCR text from entry N+1 to bleed into the fields of entry N at page column boundaries. All affected fields were truncated at the first appearance of a category marker (n.f ... pt., v. pt., adj. pt., etc., including OCR variants: y., adn., nn.fpl., hidr., inter;. and others). The truncation preserved only the content before the bleed point.

The raw_sense field in every row retains the full original OCR text before any truncation, including any bled content. This can be used to re-parse or manually recover information where truncation was too aggressive.

Entries with empty es_gloss (223 entries, 13.3%)

223 entries have no Spanish gloss. Causes:

• —Bleed truncation wiped the entire field (most common): the es_gloss field was entirely composed of the next entry's text; after truncation nothing remained.
• —Not present in the original dictionary: some entries carry only a Portuguese gloss.
• —Proper nouns (anthroponyms, toponyms): 17 of the 223 are antr/top entries where the Spanish form is identical to the Barranquenho form.

Truncated glosses at column boundaries (6 entries)

These entries have a gloss that ends at a column boundary mid-phrase. The truncated content was not recoverable from the OCR. The raw_sense field shows the available OCR text.

Verb headwords possibly missing tonic diacritic (16 entries)

The 2025 Convenção Ortográfica do Barranquenho specifies that all -ar verb infinitives end in -á (e.g. cantá, falá). The following 16 verb headwords were returned by the OCR without this accent, and the accent was not added because the raw OCR is reproduced faithfully and we cannot confirm without access to the print source whether these are genuine dictionary forms or OCR diacritic drops.

Users working on BQ NLP tasks who apply the orthographic convention programmatically should consider adding -á to these headwords where the pt_gloss confirms an -ar/-ear verb.

arrepenti-si, ehbofetia, ehfatacha, ehfranga, ehgarabata, ehgulhipa, ehmenga, ehmola, encarrilha, encerra, enchoca, entorna, entorta, enturrulha, enxamia, garri

Corrections that were applied then reverted

During validation, 26 changes were applied and subsequently reverted after identifying them as content inference rather than OCR correction:

• —12 verb headword accents added based on BQ morphological convention (e.g. ehbofetia → ehbofetiá) — reverted because the accent is not confirmed in the OCR output.
• —8 truncated gloss completions where missing words were inferred from context (e.g. taramocu es: pañuelo atado en la cabeza, kodak es: máquina de fotos) — reverted; added words were not in the source OCR.
• —3 semantic corrections: chinoti ptgloss `pequeño` was changed to `pequeno` (reverted — the source has `pequeño`), `ehmola` category changed from `v` to `n.f` (reverted — source says `v.`), `buenu` senses had esgloss cross-populated (reverted to per-sense OCR values).

The raw_sense column is the authoritative record of what the OCR returned for each entry.

Known limitations

• —Two-column boundary truncation. Approximately 30–50 entries that begin at the bottom of one column and continue at the top of the next will have their gloss or metadata fields truncated. The raw_sense field contains the available text; the missing continuation is in the adjacent column of the same page.
• —Page 84 right-column photo. A full-page botanical photograph occupies most of the right column of dictionary page 57 (PDF page 84). The three right-column entries on that page (ipidimia, ipu, íri) were recovered from the partial OCR but may have degraded metadata.
• —OCR confidence. Tesseract was trained on standard Portuguese and Spanish; Barranquenho orthography introduces characters and patterns that may reduce recognition accuracy (e.g. ehtá, uh, nóh, diacritics on final vowels like abó, cantá).
• —`raw_sense` is normalised OCR, not clean text. It contains the parser's normalisation substitutions and may contain minor residual artefacts. It is provided for auditability and re-parsing, not as a polished text field.
• —No IPA. The dictionary does not provide phonetic transcriptions; none are present in this dataset.
• —`uso` sparsity. 119 entries carry usage labels, extracted from inline {Uso: ...} tags in the original dictionary text. Remaining entries are unlabelled regardless of their actual currency.

Companion resources

The Dicionário de Barranquenho is part of a trio of 2025 codification works:

• —Gramática Básica de Barranquenho — Gonçalves, Navas, Correia (ISBN 978-972-778-464-6): full morphological paradigms, syntax, orthography rules.
• —Convenção Ortográfica do Barranquenho — Gonçalves, Navas, Correia (ISBN 978-972-778-463-9): definitive orthographic standard.

All three are published by Universidade de Évora under the same FCT-funded project.

Citation

If you use this dataset, please cite the original dictionary:

@book{goncalves2025dicionario,
  author    = {Gonçalves, Maria Filomena and Navas, María Victoria and Ferreira, Vera},
  title     = {Dicionário de Barranquenho},
  year      = {2025},
  edition   = {1},
  publisher = {Universidade de Évora},
  address   = {Évora},
  isbn      = {978-972-778-460-8}
}

And the dataset itself:

@dataset{barranquenho_dictionary_2025,
  title     = {Dicionário de Barranquenho — Structured Lexical Dataset},
  year      = {2025},
  note      = {Derived from Gonçalves, Navas \& Ferreira (2025) via OCR pipeline.
               HuggingFace Datasets.},
  language  = {Barranquenho, Portuguese, Spanish}
}

Acknowledgements

This dataset was created in support of language technology development for Barranquenho, including automatic speech recognition, text-to-speech synthesis, and phonemisation tools, as part of ongoing work on endangered and minority language NLP. The original dictionary was produced with the support of the Câmara Municipal de Barrancos and funding from the Fundação para a Ciência e a Tecnologia (FCT), project UIDB/00057/2025.