Skip to content

Latest commit

 

History

History
377 lines (249 loc) · 12.9 KB

readme.md

File metadata and controls

377 lines (249 loc) · 12.9 KB
Raw

Generalized edit distance command line tool

GenEditDist is a command line tool for finding approximate occurrences of a string in a dictionary. It employs regular edit distance (the Levenshtein distance) in combination with an user-defined set of weighted transformations to find approximate matches.

The tool has two working modes:

  • Find all the matches that are within the maximum edit distance limit;
  • Find top N closest matches;

Currently, the tool has been developed and tested only on UNIX platform and it is not expected to work on other platforms.

1. Environment and input files

The tool expects that all its inputs are in UTF-8 encoding. Before using the tool, you need to have the locales-all package installed. You can install it with the command:

apt-get install locales-all

Then you need to configure the environment to support UTF-8 on the command line. You can do this by the following commands:

export LANG=et_EE.utf8
export C_ALL=et_EE.utf8

An input 'dictionary' where the search is performed should be a text file, where each entry (a match candidate) is on a new line.

The transformations file should also be a text file, where each transformation is on a separate line. Transformation from string A to string B with cost W should be formatted in file as:

A:B:W

A or B (but not both together) can be omitted to define an addition or a deletion.

Transformations file should only contain transformations in the given format, adding empty lines or comments might result in unexpected search behaviour.

2. Using the program

2.1. Full vs partial match

While calculating the generalized edit distance, one can specify whether a full match or a partial match between the search string and a string in the dictionary is found.

Four modes of matching exist:

  1. Full extent match (flag -f, default):

    Distance between the search string and a string in the dictionary is calculated considering both strings at full extent.

    Examples (**):

    search string: police

    dictionary string: polis

    full match distance: 0.9 ( using transformation ce:s:0.9 )

    search string: computer

    dictionary string: komputa

    full match distance: 2.4 ( using transformations c:k:0.5 r::0.9 e:a:1.0 )

  2. Distance from some prefix of the dictionary word (flag -p):

    Distance between the search string and a string in the dictionary is calculated considering only a prefix of the dictionary word.

    Examples (**):

    search string: police

    dictionary string: polisman

    prefix match distance: 0.9 ( using transformation ce:s:0.9 )

    search string: book

    dictionary string: buksop

    prefix match distance: 0.5 ( using transformation oo:u:0.5 )

  3. Distance from some suffix of the dictionary word (flag -s):

    Distance between the search string and a string in the dictionary is calculated considering only a suffix of the dictionary word.

    Examples (**):

    search string: block

    dictionary string: aisblok

    suffix match distance: 0.5 ( using transformation ck:k:0.5 )

    search string: card

    dictionary string: postkat

    suffix match distance: 1.9 ( using transformations c:k:0.5 r::0.9 d:t:0.5 )

  4. Distance from some infix of the dictionary word (flag -i):

    Distance between the search string and a string in the dictionary is calculated considering only an infix of the dictionary word.

    Examples (**):

    search string: belong

    dictionary string: glas bilong kapten

    infix match distance: 0.8 ( using transformation e:i:0.8 )

    search string: something

    dictionary string: i gat samting gut

    infix match distance: 2.2 ( using transformations th:t:0.6 o:a:0.6 e::1.0 )

These matching modes can be combined, e.g flag -fi states that both full match and infix match are calculated.

2.2. Finding all matches within maximum distance

Parameter -m <max_edit_distance> is used to find all the matches that are within given maximum edit distance limit (inclusively). Usage:

./genEditDist -m <max_edit_distance> -[fpsile] <transfFile> <searchString> <dict> <caseFile>

<max_edit_distance> == maximum edit distance limit;
<transfFile> == file of user-defined transformations;
<searchString> == the search string;
<dict> == the dictionary file where the search is performed;
<caseFile> == file containing upper-to-lowercase mappings
             (optional, makes search case-insensitive)

NB! If there are several edit distances computed for a single match candidate (more than one of the flags -f, -p, -s, -i is set), then the candidate will be output if at least one of the distances is less than or equal to <max_edit_distance>

Usage examples (**):

Searching for string 'book' in the dictionary 'testdata/pidgin_words.txt' using the transformations from file 'testdata/transformations.txt':

./genEditDist  -m 1.0  -fi  testdata/transformations.txt book testdata/pidgin_words.txt
------------------------
buk
0.500000 0.500000
------------------------
buksop
3.500000 0.500000
------------------------
aisblok
4.000000 1.000000

Searching for string 'shop':

./genEditDist  -m 1.0  -fs  testdata/transformations.txt shop testdata/pidgin_words.txt
------------------------
buksop
3.600000 0.600000
------------------------
hop
1.000000 1.000000
------------------------
sop
0.600000 0.600000
------------------------
woksop
3.600000 0.600000

2.3. Finding TOP N of closest matches

Parameter -b <N> is used to find TOP N matches that are closest to the search string. Usage:

./genEditDist -b <number_of_best_matches> -[f|p|s|i] <transfFile> <searchString> <dict> <caseFile>

<number_of_best_matches> == number of closest matches displayed;
<transfFile> == file of user-defined transformations;
<searchString> == the search string;
<dictionary> == the dictionary file where the search is performed;
<caseFile> == file containing upper-to-lowercase mappings
                    (optional, makes search case-insensitive)

NB! The number of best matches is allowed to be exceeded when there are equal-distance strings for the last place in top.

Usage examples (**):

./genEditDist  -b 3 -fi  testdata/transformations.txt book testdata/pidgin_words.txt
------------------------
0.500000
buk
buksop
------------------------
1.000000
aisblok

,

./genEditDist  -b 5  -s  testdata/transformations.txt shop testdata/pidgin_words.txt
------------------------
0.600000
buksop
sop
woksop
------------------------
1.000000
hop
------------------------
1.200000
sup
sausap

2.4. Using blocked regions in the search string

Blocked regions can be used to constrain the search and allow transformations only in certain parts of the search string;

When flag -e is set, substrings surrounded with either () and <> in the search string will be blocked from changes:

  • () will block only regular edit distance changes, but allows changes through generalized edit distance transformations.

  • <> will block both regular edit distance and generalized edit distance, so the substring region cannot be changed at all.

(( or << at the beginning of the search string block additions before the first letter of the string (otherwise, such additions are allowed). Analogously, additions after the last letter are blocked by )) or >>.

Currently, additions between two consecutive regions are also blocked, so, in case of search string <a><b>c, no addition is allowed between a and b.

Usage examples (**):

./genEditDist  -m 3.0  -s -e  testdata/transformations.txt "<m>e<t>re" testdata/pidgin_words.txt
------------------------
sentimita
2.700000
------------------------
kilomita
2.700000
------------------------
milimita
2.700000
------------------------
semitrela
2.800000
------------------------
tomato
2.900000

Without using the blocked regions, the previous query (suffix matches with the word metre) gives 32 matches;

NB! If the "TOP N matches" search mode is used (flag "-b"), current implementation does not quarantee, that matches with changes in blocked regions will be left out. This is because a change in a blocked region has a concrete cost - 3000.0.

For example, if we search for string <<h>u<t> in the dictionary 'testdata/pidgin_words.txt' using the transformations from the file 'testdata/transformations.txt', only 2 matches can be reached without changing blocked regions: hat and het. So, if user queries for more than 2 closest matches, he/she will get also matches with changes in the blocked regions:

./genEditDist  -b 3  -s -e  testdata/transformations.txt "<<h>u<t>" testdata/pidgin_words.txt
------------------------
0.700000
hat
------------------------
1.000000
het
------------------------
3001.000000
i gat samting nogut
huk

2.5. Showing transformations / alignments

If the plain full extent match mode is used for finding all matches within the given maximum distance, i.e. the combination of flags -m and -f is used, and none of the flags -p, -s, -i, -e is set, then the flag -a can be used to switch on the mode of showing the transformations / alignments between the search string and each found match.

Usage examples (**):

./genEditDist -m 1.5 -f -a -y  testdata/transformations.txt  shop  testdata/pidgin_words.txt
------------------------
hop
1.000000
s:h:o:p:
 :h:o:p:
;
------------------------
sop
0.600000
sh:o:p:
 s:o:p:
;
------------------------
sup
1.200000
sh:o:p:
 s:u:p:
;

Note that the alignments are displayed not character-wise, but transformation-wise, considering both the regular edit distance transformations (single character transformations) and generalized edit distance transformations (which can also be string to string transformations). The character : separates different positions in the alignment.

If the optional flag -y is used in the mode of showing the transformations (like in the previous example), then alignments are output in a pretty-printing mode: if two aligned positions contain different length strings, the shorter string is padded with spaces from the left side.

There can be more than one best alignment between the two strings, in that case, a semicolon in a new line separates different alignments:

./genEditDist -m 3.5 -f -ay  testdata/transformations.txt shopper testdata/pidgin_words.txt
------------------------
sops
3.500000
sh:o:p:p:e:r:
 s:o:p:s: : :
;
sh:o:p:p:e:r:
 s:o:p: :s: :
;
sh:o:p:p:e:r:
 s:o: :p:s: :
;
------------------------
sop
3.500000
sh:o:p:p:e:r:
 s:o:p: : : :
;
sh:o:p:p:e:r:
 s:o: :p: : :
;

The optional flag -w can be used to show the weights / costs of the transformations. The weights are displayed in-between the two aligned strings, and colons are used as separators of different weights:

./genEditDist -m 1.5 -f -ay -w testdata/transformations.txt shop testdata/pidgin_words.txt
------------------------
hop
1.000000
s:h:o:p:
1.000000:0.000000:0.000000:0.000000:
 :h:o:p:
;
------------------------
sop
0.600000
sh:o:p:
0.600000:0.000000:0.000000:
 s:o:p:
;
------------------------
sup
1.200000
sh:o:p:
0.600000:0.600000:0.000000:
 s:u:p:
;

NB! The current implementation of showing transformations does not support partial matches (flags -p, -s, -i), finding Top N matches (flag -b) and using the blocked regions within the search string (flag -e).

3. Compiling the program

In order to compile the tool, GNU C Compiler (gcc) is needed. If the GNU 'make' utility is available, automatic compiling can be done with command:

> make all

NB! When compiling on a Solaris machine, one should make sure that GNU make is used instead of Sun's make. Usually, this can be done by calling GNU make with full path, for example /usr/sfw/bin/gmake .

4. (**) About the examples

The dictionary "testdata/pidgin_words.txt" contains words and phrases from a Pidgin English and the dictionary "testdata/english_words.txt" contains corresponding English matches. The data is based on an online Pidgin/English Dictionary (Pidgin as spoken in Port Moresby, Papua New Guinea) which can be accessed on:

http://www.june29.com/hlp/lang/pidgin.html

5. Reference

Orasmaa, S., Käärik, R., Vilo, J., & Hennoste, T. (2010). Information Retrieval of Word Form Variants in Spoken Language Corpora Using Generalized Edit Distance. In LREC.