syntok

(a.k.a. segtok v2)

Sentence segmentation and word tokenization

The syntok package provides two modules, syntok.segmenter and syntok.tokenizer. The tokenizer provides functionality for splitting (Indo-European) text into words and symbols (collectively called tokens). The segmenter provides functionality for splitting (Indo-European) token streams (from the tokenizer) into sentences and for pre-processing documents by splitting them into paragraphs. Both modules can also be used from the command-line to split either a given text file (argument) or by reading from STDIN. While other Indo-European languages could work, it has only been designed with the languages Spanish, English, and German in mind (the author's main languages).

segtok

Syntok is the successor of an earlier, very similar tool, segtok, but has evolved significantly in terms of providing better segmentation and tokenization performance and throughput (syntok can segment documents at a rate of about 100k tokens per second without problems). For example, if a sentence terminal marker is not followed by a spacing character, segtok is unable to detect that as a terminal marker, while syntok has no problem segmenting that case (as it uses tokenization first, and does segmentation afterwards). In fact, I feel confident enough to just boldly claim syntok is the world's best sentence segmenter for at least English, Spanish, and German.

Installation

To use this package, you minimally should have Python 3.6 or installed. As it uses the typing package, earlier versions are not supported. The easiest way to get syntok installed is using pip or any other package manager that works with PyPI:

pip3 install syntok

Important: If you are on Linux and have problems installing the regex dependency of syntok, make sure you have the python-dev and/or python3-dev packages installed to get the necessary headers to compile that package.

Then try the command line tools on some plain-text files (e.g., this README) to see if syntok meets your needs:

python3 -m syntok.segmenter README.rst
python3 -m syntok.tokenizer README.rst

Development

syntok uses poetry as the build tool, and expects pyenv to provide the Python versions to test with tox. Therefore, to develop syntok, it is recommended that you install Poetry and pyenv, first. Install the Python versions defined in tox.ini (see envlist) with pyenv, and set all of them as your local versions, for example: pyenv local 3.6.15 3.9.9

To run the full test suite, you have to install flake8, pytest, and mypy (via poetry install). The tests in the proper Python version & environment are run via tox or by calling the three commands by hand:

poetry run tox

# OR, manually:
poetry shell
flake8 syntok
mypy syntok
pytest syntok

Usage

For details, please refer to the code documentation; This README only provides an overview of the provided functionality.

Command-line

After installing the package, two command-line usages will be available, python -m syntok.segmenter and python -m syntok.tokenizer. Each takes [UTF-8 encoded] plain-text files (or STDIN until EOF (CTRL-D)) as input and transforms that into newline-separated sentences or space-separated tokens, respectively. You can control Python3's file open encoding by configuring the environment variable PYTHONIOENCODING to your needs (e.g. export PYTHONIOENCODING="utf-16-be"). The tokenizer produces single-space separated tokens for each input line. The segmenter produces line-segmented sentences for each input file (or after STDIN closes).

syntok.tokenizer

This module provides the Tokenizer class to tokenize input text into words and symbols (value Tokens), prefixed with (possibly empty) spacing strings, while recording their offset positions. The Tokenizer comes with utility static functions, to join hyphenated words across line-breaks, and to reproduce the original string from a sequence of tokens. The Tokenizer considers camelCase words as individual tokens (here: camel and Case) and by default considers underscores and Unicode hyphens inside words as spacing characters (not Token values). It does not split numeric tokens (without letters) if they contain symbols (e.g. maintaining "2018-11-11", "12:30:21", "1_000_000", "1,000.00", or "1..3" all as single tokens) Finally, as it splits English negation contractions (such as "don't") into their root and "not" (here: do and not), it can be configured to refrain from replacing this special "n't" token with "not", and instead emit the actual "n't" value.

To track the spacing and offset of tokens, the module contains the Token class, which is a str wrapper class where the token value itself is available from the value property and adding a spacing and a offset property that will hold the spacing prefix and the offset position of the token, respectively.

Basic example:

from syntok.tokenizer import Tokenizer

document = open('README.rst').read()
tok = Tokenizer()  # optional: keep "n't" contractions and "-", "_" inside words as tokens

for token in tok.tokenize(document):
    print(repr(token))

syntok.segmenter

This module provides several functions to segment documents into iterators over paragraphs, sentences, and tokens (functions analyze and process) or simply sentences and tokens (functions split and segment). The analytic segmenter can even keep track of the original offset of each token in the document while processing (but does not join hyphen-separated words across line-breaks). All segmenter functions accept arbitrary Token streams as input (typically as generated by the Tokenizer.tokenize method). Due to how syntok.tokenizer.Token objects "work", it is possible to establish the exact sentence content (with the original spacing between the tokens). The pre-processing functions and paragraph-based segmentation splits paragraphs, i.e., chunks of text separated by at least two consecutive linebreaks (\\r?\\n).

Basic example:

import syntok.segmenter as segmenter

document = open('README.rst').read()

# choose the segmentation function you need/prefer

for paragraph in segmenter.process(document):
    for sentence in paragraph:
        for token in sentence:
            # roughly reproduce the input,
            # except for hyphenated word-breaks
            # and replacing "n't" contractions with "not",
            # separating tokens by single spaces
            print(token.value, end=' ')
        print()  # print one sentence per line
    print()  # separate paragraphs with newlines

for paragraph in segmenter.analyze(document):
    for sentence in paragraph:
        for token in sentence:
            # exactly reproduce the input
            # and do not remove "imperfections"
            print(token.spacing, token.value, sep='', end='')
    print("\n")  # reinsert paragraph separators

Legal

License: MIT

Copyright (c) 2017-2022, Florian Leitner. All rights reserved.

Contributors

• Arjen P. de Vries, @arjenpdevries, http://www.cs.ru.nl/~arjen/
• Bastian Zimmermann, @BastianZim
• Péter Láng, @peter-lang-dealogic
• Koen Dercksen, @KDercksen, https://koendercksen.com/
• Sergiusz Bleja, @svenski

Thank you!

History

• 1.4.4 bug fixes: support for single letter consontant abbreviations, "Min.", and "Sen." #26 and German weekday abbreviations
• 1.4.3 bug fixes: under-splitting at month abbreviations #22 and over-splitting at "no." abbreviations #21
• 1.4.2 improved handling of parenthesis at start of sentences and bugfix for citations at end of texts #19
• 1.4.1 support citations at sentence begin (e.g., Bible quotes) #12
• 1.4.0 migrated to pyproject.toml and tox.ini, dropped Makefile builds and Py3.5 support
• 1.3.3 splitting tokens around the zero-width space characater U+200B #18
• 1.3.2 bugfix for offset of not contractions; discussion in Issue #15
• 1.3.1 segmenting now occurs at semi-colons, too; discussion in Issue #9
• 1.2.2 bugfix for offsets in multi-nonword prefix tokens; Issue #6
• 1.2.1 added a generic rule for catching more uncommon uses of "." without space suffix as abbreviation marker
• 1.2.0 added support for skipping and handling text in brackets (e.g., citations)
• 1.1.1 fixed non-trivial segmentation in sci. text and refactored splitting logic to one place only
• 1.1.0 added support for ellipses (back - from segtok) in
• 1.0.2 hyphen joining only should happen when letters are present; squash escape warnings
• 1.0.1 fixing segmenter.analyze to preserve "n't" contractions, and improved the README and Tokenizer constructor API
• 1.0.0 initial release