Алексей Суриков
and 1 contributors

NAME

Lingua::RU::OpenCorpora::Tokenizer - tokenizer for OpenCorpora project

SYNOPSIS

    my $tokens = $tokenizer->tokens($text);

    my $bounds = $tokenizer->tokens_bounds($text);

DESCRIPTION

This module tokenizes input texts in Russian language.

Note that it uses probabilistic algorithm rather than trying to parse the language. It also uses some pre-calculated data freely provided by OpenCorpora project.

NOTE: OpenCorpora periodically provides updates for this data. Checkout opencorpora-update-tokenizer script that comes with this distribution.

The algorithm is this:

1. Split text into chars.
2. Iterate over the chars from left to right.
3. For every char get its context (see CONTEXT).
4. Find probability for the context in vectors file (see "VECTORS FILE") or use the default value - 0.5.

CONTEXT

In terms of this module context is just a binary vector, currently consisting of 17 elements. It's calculated for every character of the text, then it gets converted to decimal representation and then it's checked against "VECTORS FILE". Every element is a result of a simple function like _is_latin, _is_digit, _is_bracket and etc. applied to the input character and few characters around it.

VECTORS FILE

Contains a list of vectors with probability values showing the chance that given vector is a token boundary.

Built by OpenCorpora project from semi-automatically annotated corpus.

HYPHENS FILE

Contains a list of hyphenated Russian words. Used in vectors calculations.

Built by OpenCorpora project from semi-automatically annotated corpus.

EXCEPTIONS FILE

Contains a list of char sequences that are not subjects to tokenizing.

Built by OpenCorpora project from semi-automatically annotated corpus.

PREFIXES FILE

Contains a list of common prefixes for decompound words.

Built by OpenCorpora project from semi-automatically annotated corpus.

NOTE: all files are stored as GZip archives and are not supposed to be edited manually.

METHODS

new(%args)

Constructs and initializes new tokenizer object.

Arguments are:

data_dir

Path to a directory with OpenCorpora data. Optional. Defaults to distribution directory (see File::ShareDir).

tokens($text [, $options])

Takes text as input and splits it into tokens. Returns a reference to an array of tokens.

You can also pass a hashref with options as a second argument. Current options:

threshold

Minimal probability value for tokens boundary. Boundaries with lower probability are excluded from consideration.

Default value is 1, which makes tokenizer do splitting only when it's confident.

tokens_bounds($text)

Takes text as input and finds bounds of tokens in the text. It doesn't split the text into tokens, it just marks where tokens could be.

Returns an arrayref of arrayrefs. Inner arrayref consists of two elements: boundary position in text and probability.

SEE ALSO

Lingua::RU::OpenCorpora::Tokenizer::Updater

http://mathlingvo.ru/nlpseminar/archive/s_49

AUTHOR

OpenCorpora.org team http://opencorpora.org

LICENSE

This program is free software, you can redistribute it under the same terms as Perl itself.