Skip to main content
Pre-release

This release is a pre-release and may not be stable for production use.

desacordo_ortografico

Detect and convert between Portuguese orthographies: the pre-1911 etymological writing, the Reforma de 1911, the Brazilian (1943 / 1971) and European (1945 / 1973) standards, and the Acordo Ortográfico de 1990, including the European and Brazilian sub-variants that AO1990 deliberately keeps distinct.

This is a single-purpose library. It does orthography detection and conversion, and nothing else.

  • Convert any text between eight orthographic norms along a graph of historical reforms, with AO1990 dual-form (facto/fato) handling.
  • Detect which orthography a text is written in: an explainable rule scorer, or two shipped zero-dependency learned classifiers (Naive Bayes / averaged perceptron).
  • Guard against sister varieties (Mirandese, Galician, Barranquenho) that look like Portuguese but are not.
  • A deterministic rule engine and curated, sourced exception lexicons. The AO1990 word maps and the pre-1911 archaisms are rows of one spelling table, data/spelling.json, built from the tugalex lists. The same table holds the 1911 words that no rule of the 1945 or 1943 edge reaches (quási gives quase), and the 1911 words that the backward pass of those edges restores (cena gives scena). Each of these rows quotes its clause. For pt-BR, data/volp_ao1990_br.json holds 1,796 spelling decisions, and each row cites its source. 1,628 rows quote the ABL Vocabulário Ortográfico da Língua Portuguesa (VOLP), 66 rows cite Priberam, and 102 rows record that no source decides the word.
  • A command-line tool and a parallel silver benchmark corpus (written by language models, not checked by a person), also published on the Hugging Face Hub.

Why it is not a single find-and-replace

Portuguese spelling is a (variant × era) matrix, not one timeline. Two official standards ran in parallel for most of the 20th century:

European / PALOP:  etymological → 1911 → 1945 → 1973 → AO1990 (PT)
Brazilian:         etymological → 1911 → 1943 → 1971 → AO1990 (BR)

AO1990 still admits divergent European/Brazilian spellings (facto/fato, António/Antônio, receção/recepção). A conversion is therefore a path through a graph of norms, applied edge by edge.

Some edges are clean regex rules (ph→f, trema abolition, the accent drops). Others are irreducibly lexical: which silent consonant drops, which form a norm prefers. Those are driven by curated, sourced data.

Install

pip install desacordo_ortografico    # pulls in tugalex

Use

from desacordo_ortografico import detect, convert, OrthographyConverter

# --- detection -----------------------------------------------------------
detect("a pharmacia do theatro").id          # 'etymological'
detect("Tambêm sei!").id                      # 'reforma_1911'  (1911 spelling)
detect("o facto é óptimo").id                 # 'pt_1973'   (pre-AO1990, European)
detect("a idéia do vôo").id                   # 'br_1971'   (pre-AO1990, Brazilian)
detect("o Antônio é econômico").id            # 'ao1990-br'

# --- conversion ----------------------------------------------------------
convert("pharmacia", "etymological", "ao1990-br")          # 'farmácia'
convert("a acção directa", "pre-ao1990-pt", "ao1990-pt")   # 'a ação direta'
convert("freqüência", "pre-ao1990-br", "ao1990-br")        # 'frequência'
convert("húmido", "ao1990-pt", "ao1990-br")                # 'úmido'
convert("Tambêm êle quási sabia", "reforma_1911", "pt_1945")  # 'Também ele quase sabia'
convert("Tambêm êle quási sabia", "reforma_1911", "br_1943")  # 'Também êle quase sabia'
convert("a pharmacia de Hollywood", "etymological", "pt_1973")
# 'a farmácia de Hollywood'  (the 1911 rules leave a foreign name as written)

# --- the full result: warnings + dual-form alternatives ------------------
conv = OrthographyConverter()
res = conv.convert("o facto é óptimo", "ao1990-pt", "ao1990-br")
res.text            # 'o facto é óptimo'  (the VOLP lists both forms of each word)
res.alternatives    # {'facto': ['facto', 'fato']}
conv.permitted_spellings("receção", "ao1990-pt")   # ['receção', 'recepção']

Names toward pt-BR

AO1990 Base XXI lets a person, a firm or a brand keep the spelling of its name. Toward ao1990-br, a word inside a sentence that starts with a capital, and is not in capitals throughout, is a name. A name keeps a form that the VOLP lists. The dual rule never gives a name a c or p that the source did not write:

convert("o setor e a Setor Norte", "ao1990-pt", "ao1990-br")
# 'o sector e a Setor Norte'
convert("o coletivo e o Coletivo Farra", "ao1990-pt", "ao1990-br")
# 'o colectivo e o Coletivo Farra'

The first word of a sentence and a text with no lowercase word give no case signal, so they convert like any other word. So does a word in capitals: o SETOR e o setor gives o SECTOR e o sector. See docs/api.md for the volp feature.

Detection: rules or a learned model

detect() defaults to an explainable rule scorer. Two zero-dependency learned models also ship: Naive Bayes and an averaged perceptron over character n-grams, sharing one sparse-dot-product scoring rule. They score higher on sentence-length text. On held-out 5-fold cross-validation over the 6,645 renderings of the silver corpus, the compatible accuracy is 96.4% for Naive Bayes, 95.6% for the perceptron and 92.6% for the rules (benchmark/train_detector.py):

detect("o Antônio era um génio econômico", method="nb")          # learned
detect("o Antônio era um génio econômico", method="perceptron")  # learned
detect("o Antônio era um génio econômico")                       # rules (default)

The guard runs first in every mode. Rule markers attach to a learned result for explainability. The learned models have no reforma_1911 class, so only the rules return that label.

A rules confidence is the held-out precision of the label for one, or two or more, evidence tokens, with two exceptions. The reforma_1911 values are provisional copies of the etymological values. An ao1990 result with a PT/BR lean marker and no AO1990 form gets 0.86: the Antônio example above returns ao1990-br with 0.86. A text with no evidence gets 0.34. A learned confidence comes from the model margin. docs/api.md gives the table and its sources.

Sister-language guard

Mirandese, Galician (especially reintegrationist), and Barranquenho look like Portuguese but are not. detect() recognizes and flags them rather than mangling them:

from desacordo_ortografico.guard import NotPortuguese
r = detect("umha cançom da naçom")
isinstance(r, NotPortuguese)   # True
r.name, r.convertible          # ('Galician', False)

CLI

desacordo detect "a pharmacia do theatro"
desacordo convert --from pre-ao1990-pt --to ao1990-pt "a acção directa"
desacordo convert --from pre-ao1990-pt --to ao1990 --variant br "a recepção"
desacordo convert -v --from ao1990-pt --to ao1990-br "o facto"
desacordo detect --method nb "a idéia do vôo"
desacordo norms          # list the available orthographic norms

--variant picks the European or Brazilian AO1990 target: the example prints a recepção, and --variant pt prints a receção. -v prints the lossy-edge warnings and the dual-form alternatives (alt: facto -> facto | fato) to stderr. --stdin reads the text from standard input.

Norms

id reform who / when
etymological pre-1911 pseudo-etymological —
reforma_1911 Reforma Ortográfica de 1911 Portugal, 1911
br_1943 Formulário Ortográfico de 1943 Brazil (ABL)
pt_1945 Convenção Luso-Brasileira de 1945 Portugal + PALOP
br_1971 accent mini-reform (Lei 5.765) Brazil, 1971
pt_1973 accent mini-reform (DL 32/73) Portugal, 1973
ao1990-pt / ao1990-br Acordo Ortográfico de 1990 all CPLP

See docs/orthographies.md for the rule sets and sources.

Reversibility

Forward conversion (toward newer norms) is well defined. Backward conversion is, for several edges, lexical and lossy. Dropping a silent consonant or an accent discards information no rule can recover. ConversionResult.lossless and .warnings report this instead of silently guessing. Back from 1945 or 1943 to 1911, only curated table rows apply: cena gives scena and ciência gives sciência (1911 Base XI), and the rules of those edges go one way.

  • tugalex — the word list this library reads, and the source of the AO1990 and archaism rows of its spelling table.

License

Apache-2.0.

Metadata

Release files for desacordo-ortografico 0.3.4a1

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for desacordo-ortografico 0.3.4a1
File Size Uploaded Provenance Actions
desacordo_ortografico-0.3.4a1.tar.gz 1.2 MB Without provenance Details for desacordo_ortografico-0.3.4a1.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for desacordo-ortografico 0.3.4a1
File Interpreter ABI Platform Actions
desacordo_ortografico-0.3.4a1-py3-none-any.whl Python 3 none any Details for desacordo_ortografico-0.3.4a1-py3-none-any.whl

Total release size: 2.3 MB

Release files / desacordo_ortografico-0.3.4a1.tar.gz

Download URL desacordo_ortografico-0.3.4a1.tar.gz
Size 1.2 MB
Tags Source
SHA-256 checksum
How to use checksums
aa3c3619184e47e05e986a02b157067676315920c3e72fca74ff61704e972a66
BLAKE2b-256 checksum
How to use checksums
47738c5612460b494fb9ef12fb463ae5ef2029e5213b1e4ff55f221b69e0698c
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.13.16

Release files / desacordo_ortografico-0.3.4a1-py3-none-any.whl

Download URL desacordo_ortografico-0.3.4a1-py3-none-any.whl
Size 1.1 MB
Tags Python 3
SHA-256 checksum
How to use checksums
1e22731058b2d22b6132694d36cbf587506359344cb1c30d9ba5d2122cafd6b0
BLAKE2b-256 checksum
How to use checksums
b6b2c5c7ae6def1ca44eb91b2bf0512d3dd9c539e801f9309b1f715fd13bdf0b
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.13.16

Security for desacordo-ortografico 0.3.4a1

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

No provenance found

None of the files in this release were uploaded with provenance metadata.

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Release history Release notifications | RSS feed

This release
Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page