latincy-preprocess

Latin text preprocessing: U/V normalization, long-s OCR correction, diacritics stripping, and macron removal — with optional Rust acceleration and spaCy integration.

Consolidates latincy-uv and latincy-long-s into a single package.

Installation

pip install latincy-preprocess

For spaCy pipeline components:

pip install latincy-preprocess[spacy]

Quick Start

from latincy_preprocess import normalize

normalize("Gallia eft omnis diuisa in partes tres")
# 'Gallia est omnis divisa in partes tres'

Per-Normalizer Usage

U/V Normalization

Converts u-only Latin spelling to proper u/v distinction using rule-based analysis:

from latincy_preprocess import normalize_uv

normalize_uv("Arma uirumque cano")
# 'Arma virumque cano'

Rules handle digraphs (qu), trigraphs (ngu), morphological exceptions (cui, fuit), positional context (initial, intervocalic, post-consonant), and case preservation.

Long-S OCR Correction

Corrects OCR errors where historical long-s (ſ) was misread as f, using n-gram frequency analysis from Latin treebank data:

from latincy_preprocess import LongSNormalizer

normalizer = LongSNormalizer()

word, rules = normalizer.normalize_word_full("ftatua")
# ('statua', [TransformationRule(...)])

text = normalizer.normalize_text_full("funt in fundamento reipublicae ftatua")
# 'sunt in fundamento reipublicae statua'

Two-pass strategy: Pass 1 applies high-confidence rules (impossible bigrams like ft, fp, fc). Pass 2 uses 4-gram frequency disambiguation for ambiguous word-initial f- patterns.

Diacritics and Macrons

from latincy_preprocess import strip_diacritics, strip_macrons

strip_macrons("ārma")
# 'arma'

strip_diacritics("λόγος")
# 'λογος'

spaCy Integration

Three pipeline components are available as spaCy factories:

Unified Preprocessor (recommended)

Chains long-s correction → U/V normalization in the correct order:

import spacy

nlp = spacy.blank("la")
nlp.add_pipe("latin_preprocessor")

doc = nlp("Gallia eft omnis diuisa in partes tres")
doc._.preprocessed          # 'Gallia est omnis divisa in partes tres'
doc[2]._.preprocessed       # 'est'
doc[2]._.preprocessed_lemma # normalized lemma

Either normalizer can be disabled:

nlp.add_pipe("latin_preprocessor", config={"uv": False})
nlp.add_pipe("latin_preprocessor", config={"long_s": False})

Standalone Components

nlp.add_pipe("uv_normalizer")
# doc._.uv_normalized, token._.uv_normalized, token._.uv_normalized_lemma

nlp.add_pipe("long_s_normalizer")
# doc._.long_s_normalized, token._.long_s_normalized

Rust Backend

When compiled with maturin, a Rust backend provides ~3x throughput for both normalizers. The backend is selected automatically:

from latincy_preprocess import backend

backend()  # 'rust' or 'python'

The Python backend is fully functional and used as the fallback.

Accuracy

U/V Normalization

Dataset	Accuracy
Curated test set (100 sentences)	100%
UD Latin PROIEL (~21K u/v chars)	~98%
UD Latin Perseus (~18K u/v chars)	~97%

Long-S Correction

Pass 1 rules have a 0.00% false positive rate. Pass 2 disambiguation uses a protected allowlist of ~170 common Latin f- words (inline in long_s/_rules.py) plus n-gram frequency tables (JSON files in long_s/data/ngrams/).

Changelog

See CHANGELOG.md for release history.

Citation

@software{latincy_preprocess,
  title = {latincy-preprocess: Text Preprocessing for LatinCy Projects},
  author = {Burns, Patrick J.},
  year = {2026},
  url = {https://github.com/latincy/latincy-preprocess}
}

License

MIT