pt2barranquenho

pt2barranquenho translates European Portuguese text into Barranquenho with hand-written rules. It is a baseline: no parallel corpus of Barranquenho exists to train a statistical model, so the rules come from the language's two reference books.

Barranquenho is the language of Barrancos, a municipality of about 1,500 people on the Portuguese border with Spain. It has a Portuguese base with Andalusian and Extremaduran Spanish features. Portugal recognised it as cultural heritage in 2021, and its first orthographic convention, grammar and dictionary appeared in 2025.

Install and use

A demo runs in the browser at huggingface.co/spaces/TigreGotico/pt2barranquenho. The page in space/ loads the package with Pyodide, so no server runs the rules.

uv pip install -e .
echo "A minha mãe disse-me que o leite estava frio." | pt2barranquenho
# A minha mãi me dissi que a lêti ehtaba friu.

from pt2barranquenho import translate, word

translate("Ontem fomos à feira e comprámos três quilos de feijão.")
# 'Onten fomu à fêra i compramu trêh quilu de fejãu.'
word("espanhol")   # 'ehpanhó'

translate takes one sentence or line. It leaves template slots such as {name} as they are, so it can convert OVOS dialog files. word converts one lowercase word with no context.

Sources

The documentation explains how the rules work, the constructions they handle, the evaluation and how to add a rule.

The rules follow two books. Each rule in the code names the section it comes from, and rules/citations.tsv quotes that section with its page; rules/README.md explains the checks and lists the places where the two books disagree.

• Gonçalves, M. F., Navas, M. V. and Correia, V. M. D. (2025). Convenção Ortográfica do Barranquenho. Évora: Universidade de Évora. ISBN 978-972-778-463-9.
• Gonçalves, M. F., Navas, M. V. and Correia, V. M. D. (2025). Gramática Básica de Barranquenho. Évora: Universidade de Évora. ISBN 978-972-778-464-6.

What the rules do

Evaluation

Four gold files test the translator. Each has the same columns; gold/README.md describes them.

"Exact" accepts the reference or any alternative printed in the books, after lowercasing. chrF2 is corpus chrF from sacrebleu.

uv run --with sacrebleu --with-editable . python scripts/evaluate.py gold/convention_text.tsv --errors

convention_text.tsv is the best test of running text. The rules were not written from it. It was first scored at chrF2 92.1 before any fix it prompted (Roman numerals, hyphenated compounds, tamén). It is short, and its Barranquenho is the authors' own rendering of a Portuguese original.

books.tsv is a development set. The rules come from the same books, so it measures how well the rules cover the books' own examples.

The external sets test unseen speech. The sources predate the 2025 norm and use their own spelling, and their Portuguese is often a free gloss. external_respelled.tsv keeps each speaker's words and grammar and changes only the spelling, following the Convenção. An independent review checked every row against the Convenção and the Gramática (gold/external_respelled_review.tsv). The respelling follows the same norm as the rules, so its scores are higher than external.tsv. Only the differences in words and grammar are evidence the respelling did not create.

A full test needs everyday Portuguese sentences that native speakers translate into the 2025 spelling.

Limits

• Barranquenho has more variation than the books record. The books themselves disagree in places: aqui and aquí, treh and trêh, ele and eli.
• The rules choose a Barranquenho-specific word only for the 13 nouns the books print. The 2025 dictionary lists many more beside their Portuguese cognates, and context decides which one a speaker uses. Used as a word list, the dictionary lowers the scores.
• Gender changes that follow Spanish (o leite > a lêti) cover only the nouns the books name. The change reaches the determiner just before the noun, not an adjective or a possessive: a minha macieira keeps minha.
• A word keeps its spelling when it has a letter or a sequence Portuguese does not use (k, w, y, ll, th, hn, a final t), because the Gramática keeps the spelling of imported names. Initials (E.), names with an apostrophe (O'Neill) or a capital inside (McMaster), and words that start with s and a consonant (Stefan) also stay. An imported name spelled like a Portuguese word still changes: Lars gives Lar.
• gostar loses de and takes a dative pronoun before a 3rd person form (le gohta tomá). In gostava and gostaria, the 1st and 3rd persons look alike. At the start of a clause the rules choose the polite 1st person of pt-PT (Me gohtaba sabê), and after a noun the 3rd person (U mê abô le gohtaba pehcá). A 3rd person pronoun liked goes with the de. The Gramática prints guhtô, a past form, for porque gostamos dela, and the rules give nuh gohta. A 1st or 2nd person liked keeps de and the Portuguese verb (gohtu muntu de ti), because the books print no form for it.
• A preterite takes ter + participle (oji tenhu compradu) only next to hoje, este ano or a similar span of time, and only in the 1st and 3rd person singular. Other uses of the compound perfect, such as Aquí tenhu conhecidu muita genti, depend on meaning.
• Stress comes from the Portuguese accent rules. A word with irregular stress can get the wrong final vowel.
• The Portuguese spelling does not show vowel quality. A final stressed e or o before -r is closed (senhô, colhê) except in a short list of words that are open in pt-PT (mulhé, melhó, pió). The noun colher (colhé) looks like the verb, so it gets colhê. The code knows caló and capela-mó only because the books print them.
• The pronoun se and the conjunction se are told apart from the neighbouring words. A sentence fragment, or a pronoun at the start of a clause, gets si.
• The 2nd person singular keeps its -h only after tú or te, and in the irregular forms no noun shares (ehtáh, tinhah). Without the pronoun, comes and come look alike to the rules. eras, foras and estavas need tú too, because eras and foras are also nouns.
• The 2nd person plural (vós) is not translated. Portuguese from Portugal uses vocês in its place, and the tables print the vós forms with variants.
• tests/test_paradigms.py checks every person the Gramática tables print for amá, comê, partí, sê, ehtá, abê and tê, except the forms where the books disagree. tests/test_closed_classes.py does the same for the pronoun, possessive, demonstrative, relative, indefinite and numeral tables.
• -amos becomes -emu, except in the present subjunctive of the verbs in a closed list of stems (comamu, partamu). A subjunctive outside the list gets -emu.
• foi, foram + infinitive count as ir only at the start of a clause or after a subject pronoun or an adverb, because they are also forms of ser.
• No native speaker has reviewed the output.

Licence

The code is Apache-2.0. The gold rows quote short examples from the two books and cite them by page. The rights to those texts stay with their authors.