choppa

Python port of the Java segment SRX sentence segmenter by Jarek Lipski, shipped with the multi-language segment.srx segmentation rules crafted by the LanguageTool team.

In a nutshell, it splits text into sentences (but it is rule-based, so you can chop anything textual). It is the same segmentation engine and the same rules LanguageTool uses — in pure Python.

• Identical output. SrxTextIterator is byte-identical to the Java original: verified on ~136,000 segments across seven real-world corpora (Ukrainian news from three sources, three English novels — zero differences) and on 848 test cases ported from LanguageTool's 24 language test suites. See Performance and verification.
• Fast. An 11 MB corpus segments in ~5 s vs ~30 s for the Java CLI (Apple M1); about 25x faster than choppa 0.9.
• Streaming. Text can be read incrementally from a file object with a fixed-size buffer, so multi-GB corpora do not need to fit in memory.

Quick Start

pip install choppa-srx

(The import name is choppa; the bare choppa name on PyPI belongs to an unrelated project.)

Command line:

echo "Жоден сучасний електронний прилад не обходиться без мікрочипів. \
Мікрочіп, інакше кажучи, мікросхема - це набір електронних схем." | choppa -l uk_two

Python:

from choppa import SrxDocument, SrxTextIterator

document = SrxDocument(ruleset="choppa/data/srx/languagetool_segment.srx")

text = "This is a sentence. And this is another one. Prof. Smith disagrees."
for sentence in SrxTextIterator(document, "en_two", text):
    print(sentence)

The language key is matched against the SRX language maps. For the bundled LanguageTool rules use <code>_two (paragraphs end at two line breaks) or <code>_one (every line break ends a paragraph): uk_two, en_two, de_one, ...

Streaming a large file:

with open("big_corpus.txt", encoding="utf-8") as reader:
    for sentence in SrxTextIterator(document, "uk_two", reader, buffer_length=1024 * 1024):
        ...

Compatibility with the Java original

The port replicates java.util.regex.Matcher semantics on top of the regex package:

• find() advancement (zero-width matches advance by one character, adjacent matches are not skipped);
• exception-rule matching with transparent bounds (pattern.match(text, pos) sees lookbehind context before pos, exactly like Java's useTransparentBounds(true) + lookingAt());
• Java-only constructs in the rules are translated: \h/\v become \p{H}/\p{V} (added to regex on our request), the (?U) flag (Java's UNICODE_CHARACTER_CLASS) is dropped because Python patterns are Unicode-aware by default, and Java character-class intersections like [\p{L}&&[^rwn]] are compiled in the regex module's V1 mode;
• rules are compiled without MULTILINE, like in Java, so ^ only matches at the start of the text. If you want ^-anchored rules to fire at every line start (pre-1.0 choppa behavior, arguably nicer for texts with headings), opt in with SrxDocument(pattern_flags=regex.M) — the analog of segment 2.0.4's defaultPatternFlags option.

Two algorithms are provided:

• SrxTextIterator — the "ultimate" algorithm from the original library and the one LanguageTool uses: finds break-rule matches, then applies merged exception patterns at each candidate break. Supports streaming. Use this one.
• AccurateSrxTextIterator — the legacy "accurate" algorithm (all rules matched over the whole text, in memory), kept for completeness. It incorporates the 2022 overlapping-exception fix from upstream.

Only the SRX 2.0 format and the SAX reader are implemented (the same parser configuration LanguageTool uses). Schema validation via xmlschema is available with SrxDocument(validate_ruleset=...) or choppa --validate.

Performance and verification

Segmentation output was compared byte-for-byte against the Java original (segment-2.0.3, ultimate algorithm, the bundled LanguageTool segment.srx, identical results under JDK 17 and JDK 24) on real-world corpora. choppa = SrxTextIterator, Apple M1, Python 3.12, regex 2026.6.28. Java times include ~1 s of JVM startup and rule parsing, so the engine gap on large inputs is what matters:

corpus	language	size	segments	Java	choppa	output
Militarny news, 100k paragraphs	uk_two	11.3 MB	43,569	29.6 s	4.6 s	identical
uanews.dp.ua, 5k articles	uk_two	12.5 MB	64,063	21.3 s	9.0 s	identical
Liga.net news, 197 articles	uk_two	0.4 MB	2,647	1.1 s	0.2 s	identical
Pride and Prejudice (Gutenberg)	en_two	0.8 MB	6,720	1.0 s	0.7 s	identical
Moby-Dick (Gutenberg)	en_two	1.3 MB	10,188	1.1 s	1.0 s	identical
Sherlock Holmes (Gutenberg)	en_two	0.6 MB	6,875	0.8 s	1.2 s	identical

In total: ~136,000 segments, zero differences.

To reproduce with your own corpus (LF line endings — Java does not normalize CRLF, so make both sides read the same bytes):

# Java ground truth (CLI zip from github.com/loomchild/segment/releases)
segment-2.0.3/bin/segment -a ultimate -s choppa/data/srx/languagetool_segment.srx \
    -l uk_two -r -i corpus.txt -o java.txt -e $'\x01'

# choppa: benchmark + byte-level diff
python scripts/benchmark.py corpus.txt --language uk_two --java java.txt

Segmentation rules

choppa/data/srx/languagetool_segment.srx is an unmodified LanguageTool segment.srx snapshot — see choppa/data/srx/PROVENANCE.md for the exact upstream commit and refresh instructions. You can also pass any SRX 2.0 file of your own via SrxDocument(ruleset=...) or choppa --srx.

Documentation

• API contract — public API, parameters, streaming semantics, error behavior, stability guarantees
• Design decisions — how Java Matcher semantics map onto Python, the rule translator, ^/MULTILINE, finitization, the two algorithms, and the testing philosophy
• Project heritage — SRX, segment, LanguageTool, and the 2022–2026 story of this port
• Performance and verification — the benchmark table above, with reproduction steps

Development

git clone https://github.com/lang-uk/choppa && cd choppa
pip install -e .[test]
pytest tests/            # includes 848 LanguageTool cases for 24 languages

scripts/benchmark.py benchmarks segmentation and byte-diffs the output against the Java segment CLI; scripts/extract_lt_tests.py regenerates the LanguageTool test fixtures from a LanguageTool checkout.

Copyrights and kudos

• Python port: Dmytro Chaplynskyi, lang-uk project
• Original Java implementation and the pure-Python matcher approach: Jarek Lipski
• CLI and packaging contributions: Volodymyr Kyrylov
• Segmentation rules: Daniel Naber, Jaume Ortolà, Andriy Rysin et al — the LanguageTool team
• Special thanks to Andriy Rysin, the driving force behind the Ukrainian language in LanguageTool, and to Matthew Barnett for adding \p{H}/\p{V} to the regex package

The port is MIT-licensed, same as the original segment library. The bundled segment.srx is distributed under LanguageTool's LGPL-2.1.