String::Approx - Perl extension for approximate matching (fuzzy matching)

  use String::Approx 'amatch';

  print if amatch("foobar");

  my @matches = amatch("xyzzy", @inputs);

  my @catches = amatch("plugh", ['2'], @inputs);

String::Approx lets you match and substitute strings approximately. With this you can emulate errors: typing errorrs, speling errors, closely related vocabularies (colour color), genetic mutations (GAG ACT), abbreviations (McScot, MacScot).

NOTE: String::Approx has been designed to work with strings, not with text. In other words, when you want to compare things like text or source code, consisting of words or tokens and phrases and sentences, or expressions and statements, you should probably use some other tool than String::Approx, like for example the standard UNIX diff(1) tool, or the Algorithm::Diff module from CPAN, or if you just want the Levenshtein edit distance (explained below), the Text::Levenshtein module from CPAN. See also Text::WagnerFischer and Text::PhraseDistance.

The measure of approximateness is the Levenshtein edit distance. It is the total number of edits: insertions,

        word world

deletions,

        monkey money

and substitutions

        sun fun

required to transform a string to another string. For example, to transform lead into gold, you need three edits:

        lead gead goad gold

The edit distance of lead and gold is therefore three, or 75%.

        use String::Approx 'amatch';

        $matched     = amatch("pattern") 
        $matched     = amatch("pattern", [ modifiers ])

        $any_matched = amatch("pattern", @inputs) 
        $any_matched = amatch("pattern", [ modifiers ], @inputs)

        @match       = amatch("pattern") 
        @match       = amatch("pattern", [ modifiers ])

        @matches     = amatch("pattern", @inputs) 
        @matches     = amatch("pattern", [ modifiers ], @inputs)

Match pattern approximately. In list context return the matched @inputs. If no inputs are given, match against the $_. In scalar context return true if any of the inputs match, false if none match.

Notice that the pattern is a string. Not a regular expression. None of the regular expression notations (^, ., *, and so on) work. They are characters just like the others. Note-on-note: some limited form of regular expressionism is planned in future: for example character classes ([abc]) and any-chars (.). But that feature will be turned on by a special modifier (just a guess: r), so there should be no backward compatibility problem.

Notice also that matching is not symmetric. The inputs are matched against the pattern, not the other way round. In other words: the pattern can be a substring, a submatch, of an input element. An input element is always a superstring of the pattern.

With the modifiers you can control the amount of approximateness and certain other control variables. The modifiers are one or more strings, for example i, within a string optionally separated by whitespace. The modifiers are inside an anonymous array: the [ ] in the syntax are not notational, they really do mean [ ], for example [ i, 2 ]. [2 i] would be identical.

The implicit default approximateness is 10%, rounded up. In other words: every tenth character in the pattern may be an error, an edit. You can explicitly set the maximum approximateness by supplying a modifier like

        number
        number%

Examples: 3, 15%.

Note that CW0% is not rounded up, it is equal to CW0.

Using a similar syntax you can separately control the maximum number of insertions, deletions, and substitutions by prefixing the numbers with I, D, or S, like this:

        Inumber
        Inumber%
        Dnumber
        Dnumber%
        Snumber
        Snumber%

Examples: I2, D20%, S0.

You can ignore case (A becames equal to a and vice versa) by adding the i modifier.

For example

        [ "i 25%", "S0" ]

means ignore case, allow every fourth character to be an edit, but allow no substitutions. (See NOTES about disallowing substitutions or insertions.)

        use String::Approx 'asubstitute';

        @substituted = asubstitute("pattern", "replacement")
        @substituted = asubstitute("pattern", "replacement", @inputs) 
        @substituted = asubstitute("pattern", "replacement", [ modifiers ])
        @substituted = asubstitute("pattern", "replacement",
                                   [ modifiers ], @inputs)

Substitute approximate pattern with replacement and return as a list <copies> of @inputs, the substitutions having been made on the elements that did match the pattern. If no inputs are given, substitute in the $_. The replacement can contain magic strings $&, $`, $' that stand for the matched string, the string before it, and the string after it, respectively. All the other arguments are as in CWamatch(), plus one additional modifier, g which means substitute globally (all the matches in an element and not just the first one, as is the default).

See BAD NEWS about the unfortunate stinginess of CWasubstitute().

        use String::Approx 'aindex';

        $index   = aindex("pattern")
        @indices = aindex("pattern", @inputs)
        $index   = aindex("pattern", [ modifiers ])
        @indices = aindex("pattern", [ modifiers ], @inputs)

Like CWamatch() but returns the index/indices at which the pattern matches approximately. In list context and if CW@inputs are used, returns a list of indices, one index for each input element. If there's no approximate match, CW-1 is returned as the index.

There's also backwards-scanning CWarindex().

        use String::Approx 'aslice';

        ($index, $size)   = aslice("pattern")
        ([$i0, $s0], ...) = aslice("pattern", @inputs)
        ($index, $size)   = aslice("pattern", [ modifiers ])
        ([$i0, $s0], ...) = aslice("pattern", [ modifiers ], @inputs)

Like CWaindex() but returns also the size (length) of the match. If the match fails, returns an empty list (when matching against CW$_) or an empty anonymous list corresponding to the particular input.

Note that the size of the match will very probably be something you did not expect (such as longer than the pattern, or a negative number). This may or may not be fixed in future releases.

If the modifier

        "minimal_distance"

is used, the minimal possible edit distance is returned as the third element:

        ($index, $size, $distance) = aslice("pattern", [ modifiers ])
        ([$i0, $s0, $d0], ...)     = aslice("pattern", [ modifiers ], @inputs)

        use String::Approx 'adist';

        $dist = adist("pattern", $input);
        @dist = adist("pattern", @input);

Return the edit distance or distances between the pattern and the input or inputs. Zero edit distance means exact match. (Remember that the match can 'float' in the inputs, the match is a substring match.) If the pattern is longer than the input or inputs, the returned distance or distances is or are negative.

        use String::Approx 'adistr';

        $dist = adistr("pattern", $input);
        @dist = adistr("pattern", @inputs);

Return the relative edit distance or distances between the pattern and the input or inputs. Zero relative edit distance means exact match, one means completely different. (Remember that the match can 'float' in the inputs, the match is a substring match.) If the pattern is longer than the input or inputs, the returned distance or distances is or are negative.

You can use adist() or adistr() to sort the inputs according to their approximateness:

        my %d;
        @d{@inputs} = map { abs } adistr("pattern", @inputs);
        my @d = sort { $d{$a} <=> $d{$b} } @inputs;

Now CW@d contains the inputs, the most like CW"pattern" first.

CWString::Approx maintains a LU (least-used) cache that holds the 'matching engines' for each instance of a pattern+modifiers. The cache is intended to help the case where you match a small set of patterns against a large set of string. However, the more engines you cache the more you eat memory. If you have a lot of different patterns or if you have a lot of memory to burn, you may want to control the cache yourself. For example, allowing a larger cache consumes more memory but probably runs a little bit faster since the cache fills (and needs flushing) less often.

The cache has two parameters: max and purge. The first one is the maximum size of the cache and the second one is the cache flushing ratio: when the number of cache entries exceeds max, max times purge cache entries are flushed. The default values are 1000 and 0.75, respectively, which means that when the 1001st entry would be cached, 750 least used entries will be removed from the cache. To access the parameters you can use the calls

        $now_max = String::Approx::cache_max();
        String::Approx::cache_max($new_max);

        $now_purge = String::Approx::cache_purge();
        String::Approx::cache_purge($new_purge);

        $limit = String::Approx::cache_n_purge();

To be honest, there are actually two caches: the first one is used far the patterns with no modifiers, the second one for the patterns with pattern modifiers. Using the standard parameters you will therefore actually cache up to 2000 entries. The above calls control both caches for the same price.

To disable caching completely use

        String::Approx::cache_disable();

Note that this doesn't flush any possibly existing cache entries, to do that use

        String::Approx::cache_flush_all();

Because matching is by substrings, not by whole strings, insertions and substitutions produce often very similar results: abcde matches axbcde either by insertion or substitution of x.

The maximum edit distance is also the maximum number of edits. That is, the I2 in

        amatch("abcd", ["I2"])

is useless because the maximum edit distance is (implicitly) 1. You may have meant to say

        amatch("abcd", ["2D1S1"])

or something like that.

If you want to simulate transposes

        feet fete

you need to allow at least edit distance of two because in terms of our edit primitives a transpose is first one deletion and then one insertion.

The starting and ending positions of matching, substituting, indexing, or slicing can be changed from the beginning and end of the input(s) to some other positions by using either or both of the modifiers

        "initial_position=24"
        "final_position=42"

or the both the modifiers

        "initial_position=24"
        "position_range=10"

By setting the position_range to be zero you can limit (anchor) the operation to happen only once (if a match is possible) at the position.

Major release 3.

The version 3 is 2-3 times faster than version 2

No pattern length limitation

The algorithm is independent on the pattern length: its time complexity is O(kn), where k is the number of edits and n the length of the text (input). The preprocessing of the pattern will of course take some O(m) (m being the pattern length) time, but CWamatch() and CWasubstitute() cache the result of this preprocessing so that it is done only once per pattern.

You do need a C compiler to install the module

Perl's regular expressions are no more used; instead a faster and more scalable algorithm written in C is used. The string matched and substituted is now always stingy, as short as possible. It used to be as long as possible. This is an unfortunate change stemming from switching the matching algorithm. Example: with edit distance of two and substituting for word from cork and wool previously did match cork and wool. Now it does match or and wo. As little as possible, or, in other words, with as much approximateness, as many edits, as possible. Because there is no need to match the c of cork, it is not matched.

The following people have provided valuable test cases, documentation clarifications, and other feedback:

Jared August, Arthur Bergman, Anirvan Chatterjee, Steve A. Chervitz, Aldo Calpini, David Curiel, Teun van den Dool, Alberto Fontaneda, Rob Fugina, Dmitrij Frishman, Lars Gregersen, Kevin Greiner, B. Elijah Griffin, Mike Hanafey, Mitch Helle, Ricky Houghton, 'idallen', Helmut Jarausch, Damian Keefe, Ben Kennedy, Craig Kelley, Franz Kirsch, Dag Kristian, Mark Land, J. D. Laub, John P. Linderman, Tim Maher, Juha Muilu, Sergey Novoselov, Andy Oram, Ji Y Park, Eric Promislow, Nikolaus Rath, Stefan Ram, Slaven Rezic, Dag Kristian Rognlien, Stewart Russell, Slaven Rezic, Chris Rosin, Pasha Sadri, Ilya Sandler, Bob J.A. Schijvenaars, Ross Smith, Frank Tobin, Greg Ward, Rich Williams, Rick Wise.

The matching algorithm was developed by Udi Manber, Sun Wu, and Burra Gopal in the Department of Computer Science, University of Arizona.

Jarkko Hietaniemi <jhi@iki.fi>

Copyright 2001-2005 by Jarkko Hietaniemi

This library is free software; you can redistribute it and/or modify it under the same terms as Perl itself.

Furthermore: no warranties or obligations of any kind are given, and the separate file COPYRIGHT must be included intact in all copies and derived materials.