Sentence boundary disambiguation (SBD) - or sentence breaking - library written in PHP.

composer require bigwhoop/sentence-breaker

<?php
use Bigwhoop\SentenceBreaker\SentenceBreaker;

$breaker = new SentenceBreaker();
$breaker->addAbbreviations(['Dr', 'Prof']);

// returns a generator, the text is parsed lazily
$sentences = $breaker->split("Hello Dr. Jones! How are you? I'm fine, thanks!");

// get first
$sentences->current() // 'Hello Dr. Jones!'

// get all as array
iterator_to_array($sentences) // ['Hello Dr. Jones!', 'How are you?', "I'm fine, thanks!"]

By default the rules/rules.ini file is loaded. Its format is a list of patterns ...

TOKEN [... TOKEN] = PROBABILITY
T_CAPITALIZED_WORD <T_PERIOD> T_WHITESPACE T_CAPITALIZED_WORD = 75

The token enclosed in < / > is the one that defines for which token the pattern is applied. The example pattern above would be applied to each T_PERIOD token found in the input data. The probability defines how likely a sentence boundary is after this token.

So for this pattern to match, the input text would need to contain something along the lines of This is Waldo. He likes dogs..

The available tokens are:

TIP: You can add your own rules via $breaker->addRules().

Inside the data directory are flat files containing abbreviations (in English), collected from various sources. They can be loaded like this:

use Bigwhoop\SentenceBreaker\Abbreviations\FlatFileProvider;

// Load legal.txt and biz.txt
$breaker->addAbbreviations(new FlatFileProvider('/path/to/data/directory', ['legal', 'biz']));

// Load all files
$breaker->addAbbreviations(new FlatFileProvider('/path/to/data/directory', ['*']));

To make it fast and easy, all abbreviations are available in the all.txt file. You can load it like this:

$breaker->addAbbreviations(new FlatFileProvider('/path/to/data/directory', ['all']));

The input text is run through a lexer.

In computer science, lexical analysis is the process of converting a sequence of characters into a sequence of tokens, i.e. meaningful character strings.

So for example He asked: "What's on TV?" On T.V.? I have no clue. Really! would result in the following sequence of tokens:

"He" "asked:" T_QUOTED_STR "On" "T.V" T_PERIOD T_QUESTION_MARK
"I" "have" "no" "clue" T_PERIOD "Really" T_EXCLAMATION_POINT

This sequence of tokens is then run through a probability calculator that calculates for each token the probability of it being the boundary of a sentence. The calculator uses rules that are matched against each token. For example if a T_EXCLAMATION_POINT is followed by a capitalized string the chance of it being a sentence boundary is 100%.

In the end the tokens are re-assembled into the sentences. The user can choose which threshold he wants to apply when starting new sentences. For example the probability must be greater or equal to 50% that a boundary was detected.

• calculateCurrentTokenProbability is a big mess. Let's split it up into multiple Rule classes. Maybe use a rules engine.
• Add abbreviations support for different languages.

# Check code style
vendor/bin/php-cs-fixer fix --diff --dry-run

# Fix code style
vendor/bin/php-cs-fixer fix --diff

# Run tests
vendor/bin/phpunit

# Run static analysis
vendor/bin/phpstan

Made with contrib.rocks.

MIT. See LICENSE file.