nltk-punkt-tokenize

A modern implementation of the Punkt unsupervised sentence boundary detector (Kiss & Strunk, 2006) that does not read model files at all.

Zero dependencies. Standard library only. Pretrained models for 19 languages ship inside the wheel as compiled Python modules; nothing is downloaded, ever, and nothing is parsed.

The dependency direction is deliberate: this package is intended for NLTK to depend on, so that Punkt can be vendored out of NLTK. Nothing here imports NLTK — not the library, not the tests, not the CI.

pip install nltk-punkt-tokenize

import punkt

punkt.sent_tokenize("Dr. Smith went home. He was tired.")
# ['Dr. Smith went home.', 'He was tired.']

punkt.sent_tokenize("Das ist z.B. ein Satz. Und noch einer.", "german")
# ['Das ist z.B. ein Satz.', 'Und noch einer.']

punkt.sent_tokenize("这是一支笔。那是一本书。", "chinese")
# ['这是一支笔。', '那是一本书。']

On WMT24++, 21 of 21 languages find every segment boundary — 0 missed, 100.00% recall across Latin, Cyrillic, Greek, Chinese, Japanese and Korean. See Accuracy.

---

Why this exists

Punkt is a good algorithm with a bad delivery mechanism. NLTK's models were distributed as Python pickles fetched at run time, which is CVE-2024-39705: the pickle grammar contains an opcode meaning "import this module, look up this name, and call it with these arguments", so loading a model is arbitrary code execution. No amount of care in the calling code changes that, because the file chooses what gets called.

The usual reflex — swap pickle for json and declare victory — does not finish the job. JSON cannot name a class, so it is not a code-execution primitive; but json.loads builds the entire object graph before any validation code runs, which leaves the whole denial-of-service family intact.

The conclusion this package eventually reached is that the safest parser is the one that is not there. A Punkt model is four containers of strings and integers, which Python can already express, so the models are compiled into the package as Python literals and the library ships no model reader of any kind. Reading files is a separate job for separate code you run deliberately — see tools/ — and not something an installed library does on your behalf. See Security.

What's different

Models are data, not programs. A Punkt model is four containers of strings and integers. Every model the package can load is compiled into it as Python literals, so loading one is an import — no file to find, no format to parse, no initialisation.

The package contains no model parser at all. Not for pickles, not for JSON, not for its own format. There is nothing for a model file to attack, because there is nothing that reads one. The readers for the four file formats a Punkt model has historically been stored in live in tools/ and are deliberately not shipped: they exist to produce the compiled modules, and to bring in a model you already have.

The layers are separated. NLTK's punkt.py is one 1,800-line module where the language rules, the learner, and the tokenizer share mutable state through a common base class. Here:

Layer	Holds	Depends on
punkt.rules	Language conventions, regexes, the token object	nothing
punkt.model	Learned parameters as data, plus every codec	rules
punkt.annotate	The Kiss & Strunk decision procedure	rules, model
punkt.training	The learner. Writes models, never tokenizes	the above
punkt.inference	The tokenizer. Reads models, never trains	the above

The trainer has no tokenize(). The tokenizer has no train(). Hyperparameters live in a frozen TrainerConfig rather than as class attributes you have to monkey-patch.

It matches NLTK exactly. NLTK's output was captured once, from identical parameters, into tests/data/parity.json. The suite asserts against that snapshot: equal sentences and spans over 400 generated documents, equal trained models, and bit-exact log-likelihood scores. Because the reference is recorded rather than imported, parity is checked on every machine and in every CI job — and NLTK never has to be installed to check it. Four divergences are deliberate and asserted as such, so they cannot quietly become five.

Usage

Splitting

tokenizer = punkt.PunktSentenceTokenizer(punkt.load_model("english"))

tokenizer.tokenize(text)                          # list[str]
list(tokenizer.span_tokenize(text))               # [(start, end), ...]
tokenizer.tokenize(text, realign_boundaries=False)

Whitespace inside a sentence is preserved exactly, including newlines. Only whitespace between sentences is dropped.

Training

Punkt learns from unannotated text — no labelled sentence boundaries required.

params = punkt.train(corpus_text, language="danish")

A trained model is a value you can use immediately. To keep one, render it as a Python module — the same representation the bundled models use, so there is no second format that only your models are stored in:

Path("danish.py").write_text(punkt.render_model(params, "danish"))

from danish import MODEL
punkt.PunktSentenceTokenizer(MODEL)

Rendering only ever writes literals, and verifies its own output by parsing it and rejecting anything that is not data — see Security.

Incrementally, for a corpus too large to hold in memory:

trainer = punkt.PunktTrainer()
for chunk in chunks:
    trainer.train(chunk, finalize=False)
params = trainer.get_params()

Tuning:

config = punkt.TrainerConfig(abbrev_threshold=0.5, include_all_collocations=True)
params = punkt.train(corpus_text, config=config)

Understanding a decision

for decision in tokenizer.debug_decisions("Dr. Smith went home."):
    print(punkt.format_decision(decision))

Text: 'Dr. Smith' (at offset 2)
Sentence break? False (default decision)
Collocation? False
'dr.':
    known abbreviation: True
    is initial: False
'smith':
    known sentence starter: False
    orthographic heuristic suggests is a sentence starter? unknown
    orthographic contexts in training: {'MID-UC', 'UNK-UC'}

Documents with headlines and captions

Punkt keys entirely off punctuation, so a line that ends a sentence by layout rather than with a period is invisible to it. On WMT24++ that is where nearly every missed boundary came from:

punkt.sent_tokenize(text, line_breaks="always")     # any newline ends a sentence
punkt.sent_tokenize(text, line_breaks="paragraph")  # a blank line does

The default is "ignore" — correct for hard-wrapped prose, where newlines fall mid-sentence. Pick by how your text is formatted.

Abbreviations the model never saw

The bundled English model was trained on the Wall Street Journal, so it knows Dr. and Corp. but not e.g. or i.e. — the source of the most-reported Punkt failures (nltk#2376, #2154, #3370, all open).

punkt.sent_tokenize("Use a hammer, e.g. a claw hammer. Then hit the nail.", prefixes=True)
# ['Use a hammer, e.g. a claw hammer.', 'Then hit the nail.']

This is opt-in, because every abbreviation added is a boundary that can no longer be found: "Add water, sugar, etc. Then stir." stops splitting. Good trade for technical prose, bad for narrative — measure on your own text.

etc, al and dept are held back even from that list, because each one genuinely ends sentences as often as not. They become placeable once the model can judge the following word:

params = punkt.with_nonbreaking_prefixes(punkt.load_model("english_web"),
                                         "english", ambiguous=True)
tok = punkt.PunktSentenceTokenizer(params)
tok.tokenize("Cats, dogs, etc. are common pets.")   # 1 sentence
tok.tokenize("Add water, sugar, etc. Then stir.")   # 2 sentences

On the bundled english model the same flag is a pure trade and gains nothing; on english_web it takes those six cases from 3/6 to 5/6.

A subset is safe unconditionally. Moses marks some prefixes as non-breaking only before a number, which a flat list cannot express:

punkt.sent_tokenize("See No. 5 on the list. It is important.")
# ['See No.', '5 on the list.', 'It is important.']   <- wrong

params = punkt.with_nonbreaking_prefixes(punkt.load_model("english"), "english",
                                         include_always=False)
punkt.PunktSentenceTokenizer(params).tokenize("See No. 5 on the list. It is important.")
# ['See No. 5 on the list.', 'It is important.']      <- and "There is no. Not at all." still splits

A new language

Language rules are an immutable value, not a subclass:

greek = punkt.LanguageVars(sent_end_chars=(".", ";", "·", "!"))
tokenizer = punkt.PunktSentenceTokenizer(params, rules=greek)

Ready-made rule sets ship for the languages the defaults do not fit:

from punkt.rules.presets import CHINESE, JAPANESE, KOREAN, GREEK, THAI

Chinese and Japanese need them structurally, not cosmetically: those scripts put no space between sentences, so the default lookahead finds zero boundary candidates and returns the whole text as one sentence. Tan & Bond (2011) abandoned Punkt over exactly this when building the NTU-MC corpus.

Accuracy

Measured with tools/eval_wmt24pp.py against WMT24++ segment boundaries, using line_breaks="always" since those documents are newline-separated paragraphs:

	recall	missed boundaries
21 languages, this release	100.00%	0 of 790 each
punctuation only (line_breaks="ignore")	65–78%	172–274 each

Only recall is reported. A WMT24++ segment is a translation segment, not a sentence — roughly half the English ones hold more than one — so a split inside a segment is usually a correct sentence break that the segmentation did not record. Counting those against a splitter would punish it for being right. Missing a segment boundary, on the other hand, is unambiguously wrong.

Command line

punkt tokenize book.txt                 # one sentence per line
punkt tokenize -m german artikel.txt
punkt spans book.txt                    # start, end, sentence
cat corpus.txt | punkt train - -o mine.py -l english
punkt info english --show
punkt explain "Dr. Smith went home."
punkt languages

Migrating from NLTK

NLTK	here
nltk.sent_tokenize(text)	punkt.sent_tokenize(text)
nltk.download("punkt")	not needed — models are bundled
PunktSentenceTokenizer(train_text)	PunktSentenceTokenizer(punkt.train(train_text))
PunktTrainer.ABBREV = 0.5	TrainerConfig(abbrev_threshold=0.5)
tokenizer._params	tokenizer.params
PunktLanguageVars subclass	LanguageVars(...) value

Models you already have in nltk_data

The 19 stock languages are bundled, so you need nothing. For a model you trained yourself, or a punkt_tab directory you want to bring across, clone this repository and use the converter — it is not part of the installed package, by design:

python tools/convert_model.py --list                     # what nltk_data has
python tools/convert_model.py ~/nltk_data/tokenizers/punkt_tab/portuguese \
    -o portuguese.py

from portuguese import MODEL
punkt.PunktSentenceTokenizer(MODEL)

.pickle inputs are parsed, never unpickled. --all converts a whole directory, and --format punkt_tab writes a directory NLTK can load. See tools/README.md.

Deliberate differences from NLTK 3.8.1

• Unicode quotes. Curly quotes and guillemets are treated as closing punctuation, so “Hello there.” Bye. realigns correctly. NLTK added this after 3.8.1 (gh-1682); this package follows the newer behaviour.
• Abbreviation smoothing. The abbreviation log-likelihood adds 1e-8 to its null probability, matching NLTK's development branch. Without it a corpus containing no period-final tokens raises ValueError from log(0).
• No mutation on read. NLTK stores ortho_context in a defaultdict(int) and reads it with [], so tokenizing silently grows the model with a zero entry per unknown word. Reads here do not mutate.
• Real booleans. is_initial and friends return bool, not a truthy re.Match. The orthographic heuristic returns a three-valued enum instead of True/False/"unknown" in one variable.

Why the name

nltk-punkt on PyPI is an unrelated 1.2 KB package that downloads NLTK's punkt data — the very thing this replaces — so the distribution is nltk-punkt-tokenize. The import name is just punkt.

Security

The shape of the thing

The package does not read model files. Every model it can load is a Python module inside the wheel, so punkt.load_model is an import: there is no path to resolve, no bytes to parse, no format to confuse, and no limit to exceed. Whole categories of attack are absent rather than defended against.

There is also no search path. Earlier releases consulted $PUNKT_MODEL_PATH and ~/.punkt/models before the compiled-in data, which let a file dropped in the right directory decide what your program tokenized with. Names now resolve to bundled modules and nothing else.

The one place code runs

Importing a module executes it. That is the trust boundary every Python package already has — you trust the code you installed — and it is why the bundled route is used only for models inside the wheel, covered by the wheel's own hash and reviewed at release. What made CVE-2024-39705 a vulnerability was not that NLTK's models were code; it was that they were downloaded at run time and then unpickled. Nothing here downloads anything.

Two tests enforce it: one walks the package AST asserting no unpickling call exists, the other walks the generated data modules asserting they contain nothing but assignments.

Rendering a model as Python

punkt.render_model writes Python source that you then import, and some of what it writes can come from a model read out of a file somebody else wrote. That is a code-execution vector if any value reaches the output unquoted — and one did. The model name was interpolated into the generated module's docstring without repr, so a name containing a triple quote closed the docstring and everything after it became live code. tools/convert_model.py derives that name from the input file's name, which made a maliciously named file arbitrary code execution at import time.

Three things now stand in the way:

1. The name must be a Python identifier — the right constraint, since it becomes a module name, and a complete one.
2. Every value must be a str or an int, checked before rendering, and each is emitted through repr.
3. The finished source is parsed and verified: the AST may contain only literals, tuple and dict displays, and calls to frozenset, and must assign exactly the ten expected names. Anything else raises rather than being written.

The third is the guarantee that does not rest on the quoting being right. tests/test_render.py attacks all three through the name, the metadata and each learned container, and asserts a fresh interpreter importing the result has no side effects.

Reading other people's files

The converters under tools/ do parse files, and they carry the defences that implies. They are not shipped and the library never calls them; you run them deliberately, on input you chose.

Attack	Defence
Code execution via pickle	Nothing calls pickle.load. Pickles are parsed with pickletools.genops and replayed through a data-only machine; GLOBAL pushes an inert marker, REDUCE consults a fixed table, PERSID/EXT* are refused
Gzip bomb	Decompressed bytes counted while streaming and capped
Multi-gigabyte model	Size, line count, line length and per-section entry caps
JSON nesting exhausting the parser stack	Bracket depth counted before json.loads is called
Quadratic number parsing	Numeric literal length is bounded. int(digits) is O(n²), and Python 3.9/3.10 have no ceiling — a 2 MB run of digits passes both a depth check and a size check, then pins a core for minutes. Measured: >120 s unguarded, 0.09 s guarded
Hash/memory pressure from huge objects	Total element count is bounded before parsing. A post-hoc len() cannot help: the dict already exists
Unbounded JSON string	String literal length is bounded
Schema confusion	Every JSON field is type- and range-checked; unknown keys are refused, not ignored
Malformed data loading as a different model	Strict UTF-8, no errors="replace"; unknown sections rejected; flag bits validated against a mask
Callback-driven JSON parsing	object_hook, object_pairs_hook and parse_constant are never passed

The four JSON bounds are applied to the raw text in a single linear pass before the parser is invoked at all, because anything checked after json.loads returns is checked too late.

Performance

Loading a bundled model is an unmarshal of a precompiled .pyc rather than a parse:

	time
Parse a 20,000-entry text model	~66 ms
Import the equivalent compiled module	~5 ms

The first import of a language compiles its module to a .pyc — 200 ms for English, 800 ms for Finnish, once per installation. pip normally does this at install time, so it is not usually observed.

Training caches the corpus token total, which NLTK recomputes inside two per-type loops; this makes training linear in vocabulary size rather than quadratic, with identical results.

Models

19 languages, from Kiss & Strunk's original multilingual evaluation plus later contributions, as redistributed by NLTK: Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Italian, Malayalam, Norwegian, Polish, Portuguese, Russian, Slovene, Spanish, Swedish, Turkish. Plus english_web, described below.

punkt.available_languages()

A name resolves to a bundled module and nothing else — there is no search path and no way for a file on disk to shadow one. To use a model of your own, build it (punkt.train) or convert it (tools/convert_model.py) and import it directly.

The Russian model contains abbreviations only, with no orthographic data. That is how it is distributed upstream; it will detect fewer boundaries than the others.

english_web

One extra model, opt-in. The bundled english is the Wall Street Journal model, which knows 39 sentence starters and 20,366 orthographic types. Punkt's second pass uses exactly those to decide whether an abbreviation also ends a sentence, so with that little evidence it usually answers "unknown".

english_web keeps the same abbreviations and adds evidence from 400 MB of HPLT 2.0 web text: 908 sentence starters, 381,201 orthographic types. On Universal Dependencies English it takes punctuated F1 from 0.9785 to 0.9838.

tokenizer = punkt.PunktSentenceTokenizer(punkt.load_model("english_web"))

english is untouched, so parity with NLTK and existing output are unaffected. Regenerate the model with:

python tools/train_statistics.py english --bytes 400000000 --emit english_web

The same command works for any of 35 languages. Abbreviations learned from the crawl are dropped — crawl is good at statistics and bad at abbreviations, and 400 MB of English yielded 1,648 new ones that were almost entirely $40000, !m and 3c•. Curated abbreviations come from punkt.data.curated instead.

Requirements

Python 3.9+. No dependencies, at runtime or otherwise — the package imports only the standard library, and pip install nltk-punkt-tokenize pulls in nothing else. CI asserts this three ways: over the AST, over the distribution metadata, and by making import nltk raise and then exercising the package.

Development

git clone https://github.com/alvations/nltk_punkt
cd nltk_punkt
pip install -e ".[dev]"
pytest                    # 405 tests: the package
pytest tools/tests        # 163 tests: the converters
ruff check src tools tests
mypy src/punkt

Two suites, because there are two things. tests/ covers what the wheel contains, and must pass with tools/ absent — which is how it arrives in an sdist. tools/tests/ covers the model-file converters, which are published in no artefact at all and run only from a checkout.

Everything that builds, converts or scores lives in tools/. See the tools README for regenerating the bundled data, converting models out of nltk_data, and why writing Python from an untrusted file is safe here.

References

Kiss, T. & Strunk, J. (2006). Unsupervised Multilingual Sentence Boundary Detection. Computational Linguistics, 32(4), 485–525.

Dunning, T. (1993). Accurate Methods for the Statistics of Surprise and Coincidence. Computational Linguistics, 19(1), 61–74.

Licence

Apache-2.0. The pretrained models are redistributed from the NLTK project and were trained by Jan Strunk and Tibor Kiss; see NOTICE.