Ted Pedersen > Text-SenseClusters-1.01 > wordvec.pl

Download:
Text-SenseClusters-1.01.tar.gz

Annotate this POD

CPAN RT

Open  0
View/Report Bugs
Source   Latest Release: Text-SenseClusters-1.03

NAME ^

wordvec.pl - Construct word vectors from bigram or co-occurrence matrices

SYNOPSIS ^

 wordvec.pl [OPTIONS] WORD_PAIRS

DESCRIPTION ^

Constructs word vectors from the given WORD_PAIRS.

INPUT ^

Required Arguments:

WORD_PAIRS

WORD_PAIRS should be a bigram or co-occurrence pair file as created by programs count.pl, statistic.pl or combig.pl from the N-gram Statistics package.

Various ways to create WORD_PAIRS are -

1. Run count.pl alone
 (WORD_PAIRS show bigram frequency counts)
2. Run count.pl followed by combig.pl
 (WORD_PAIRS show co-occurrence pair frequency counts)
3. Run count.pl followed by statistic.pl
 (WORD_PAIRS show test of association scores of bigrams)
4. Run count.pl followed by combig.pl followed by statistic.pl
 (WORD_PAIRS show test of association scores of co-occurrence pairs)

Cases 1 and 2 will create WORD_PAIRS in format -

        word1<>word2<>n11 n1p np1 

where n11 shows the joint bigram or co-occurrence frequency count

Cases 3 and 4 will create WORD_PAIRS in format -

        word1<>word2<>rank score n11 n1p np1 

where 'score' shows the test of association score of a bigram/co-occurrence pair.

Optional Arguments:

--wordorder WORDORD

Allows to retain or ignore the order of the words in the WORD_PAIRS. The possible options for the value of --wordorder are -

--binary

Creates binary word vectors that show mere presence (by 1) or absence (by 0) of the feature-dimension pairs. By default, wordvec creates frequency vectors that show the frequency scores of the word pairs as given in the WORD_PAIRS file.

--dense

Creates dense word vectors. By default, output of wordvec will show sparse word vectors.

--feats FEATFILE

Specifies the name of the feature file that lists the words that represent the rows of the output word association matrix.

If the FEATFILE exists, words listed in this file define the rows of the output word matrix. Thus, the FEATFILE specifies the feature words for which the word vectors are to be created.

If the FEATFILE doesn't exist, it is created by wordvec and shows the words that represent the rows of the output word matrix.

--dims DIMFILE

DIMFILE is created by wordvec and reports the words that represent the columns/dimensions of the output word matrix.

--target TARGET_REGEX

Specifies a file containing Perl regex/s that define the target word. By default, target.regex file is assumed to exist in the current directory. This is only required if --extarget is selected.

--extarget

This will ignore WORD_PAIRS in which either of the constituent words is a target word. Target word can be defined by specifying a target regex file via --target option or by copying target.regex file to current directory.

--format FORM

Specifies numeric format for representing each word vector entry.

Possible values of FORM are

 iN   -> integer format allocating total N bytes/digits for each entry

 fN.M -> floating point format allocating total N bytes/digits for each entry of which last M digits show fractional part. 

When --binary is ON, default format is i2 and otherwise default is f16.10.

Other Options :

--help

Displays this message.

--version

Displays the version information.

OUTPUT ^

Consider the following illustration -

Sample WORD_PAIRS input =>

 stir<>soup<>5 21 64
 soup<>plate<>8 70 14
 hot<>soup<>12 173 64
 hot<>plate<>9 173 29
 salt<>pepper<>42 124 121
 taste<>salt<>18 83 84
 add<>salt<>12 157 84
 stir<>lemon<>2 21 53
 lemon<>juice<>2 10 2
 add<>lemon<>3 157 53
 lemon<>pepper<>3 67 120
 stir<>juice<>2 21 27
1. --wordorder = follow or default

Given WORD_PAIRS are treated as bigrams and the order of the words is retained such that the 1st word in the bigrams becomes a feature and is assigned a unique row index while the 2nd word becomes a dimension and is assigned a single column index in the output word matrix.

case (1) Feature file provided via --feats FEATFILE doesn't exist

Feature file is automatically created by wordvec and lists all the word types that appear as the 1st words in the given bigrams. i.e. -

 stir<>
 soup<>
 hot<>
 salt<>
 taste<>
 add<>
 lemon<>

The dimension file created with --dims option will list all the word types that appear as the 2nd words in the given bigrams. i.e. -

 soup<>
 plate<>
 pepper<>
 salt<>
 lemon<>
 juice<>

Thus, the bigrams listed in the given WORD_PAIRS file can be viewed in a matrix form as -

          soup<>  plate<>  pepper<>  salt<>  lemon<>  juice<>
 stir<>   5       0        0         0       2        2
 soup<>   0       8        0         0       0        0
 hot<>   12       9        0         0       0        0
 salt<>   0       0       42         0       0        0
 taste<>  0       0        0        18       0        0
 add<>    0       0        0        12       3        0
 lemon<>  0       0        3         0       0        2

whose rows represent the feature words and columns represent the dimension words.

a. --dense not used

By default, the output word matrix is created in sparse format in which the first line shows

 #rows #cols #nnz 

i.e. number of rows, number of columns and total number of non-zero entries separated by space.

Each line thereafter shows a sparse word vector of the feature shown on the corresponding line in the feature file.

A sparse word vector lists pairs of numbers separated by space such that the first number in a pair indicates the column index of a non-zero value and the second number is the value itself that appears at that index.

Thus, the output of wordvec for the above example, created with --wordorder = follow or un-specified will be =>

 7 6 12
 1  5 5  2 6  2
 2  8
 1 12 2  9
 3 42
 4 18
 4 12 5  3
 3  3 6  2

where the 1st line "7 6 12" shows that there are total 7 word vectors represented using 6 dimensions with 12 non-zero entries.

Each row thereafter indicates a word vector in sparse format. e.g. 2nd line shows the word vector of feature stir<>. This vector has total 3 non-zero values 5, 2, 2 that occur at indices 1, 5, 6 resp.

Column index counting starts from 1 to be consistent with Cluto's matrix format.

b. --dense used

When --dense is used, output will show the word matrix in dense format as

  7 6
  5  0  0  0  2  2
  0  8  0  0  0  0
 12  9  0  0  0  0
  0  0 42  0  0  0
  0  0  0 18  0  0
  0  0  0 12  3  0
  0  0  3  0  0  2

where the first line shows that there are 7 word vectors represented using 6 dimensions.

c. --binary used

When --binary is used, all non-zero bigram scores will be set to 1.

Thus, when --dense is used, output will show

  7 6
  1  0  0  0  1  1
  0  1  0  0  0  0
  1  1  0  0  0  0
  0  0  1  0  0  0
  0  0  0  1  0  0
  0  0  0  1  1  0
  0  0  1  0  0  1

Otherwise, binary sparse vectors will look like -

 7 6 12
 1  1 5  1 6  1
 2  1
 1  1 2  1
 3  1
 4  1
 4  1 5  1
 3  1 6  1

case (2) Feature file provided via the '--feats FEATFILE' option exists and lists the features for which the vectors are to be created.

Suppose the FEATFILE contains -

 taste<>
 hot<>
 lemon<>
 salt<>

Then, for each bigram word1<>word2, if word1 is one of the above words listed in the FEATFILE, a unique row index say i is assigned to word1 and a unique column index say j is assigned to word2. The matrix entry at [i][j] then indicates the score of the bigram word1<>word2. Thus, for the above example, the word matrix can be viewed as -

        soup  plate  pepper  salt  juice
 taste     0      0       0    18      0
 hot      12      9       0     0      0
 lemon     0      0       3     0      2
 salt      0      0      42     0      0

The dimension file created with --dims option will show the words that represent the columns -

 soup<>
 plate<>
 pepper<>
 salt<>
 juice<>

The output of wordvec created with --dense option will look as -

 4 5
 0      0       0    18      0
 12     9       0     0      0
 0      0       3     0      2
 0      0      42     0      0

where, the first line shows that there are 4 features and 5 dimensions. Each line thereafter shows the word vector of the corresponding feature word.

The following shows the sparse representation of the same matrix when --dense is not used -

 4 5 6
 4 18
 1 12 2 9
 3 3 5 2
 3 42

where the first line indicates that there are total 4 features, 5 dimensions and total 6 non-zero entries in the output matrix. Each row after that shows the 'index value' pair for each non-zero entry at that row, where column indices start with 1s.

2. --wordorder = precede

Order of words in bigram pairs is retained such that 2nd word becomes a feature and represents a row while the 1st word becomes a dimension and represents a column of the output word matrix. The word matrix thus shows the transpose of the bigram matrix created by --wordorder = follow and the cell values show how frequently a dimension word precedes a feature word.

The feature file created with --feats option shows the word types that appear as the 2nd words in the given bigrams. i.e.

 soup<>
 plate<>
 pepper<>
 salt<>
 lemon<>
 juice<>

while the dimension file created with dims option shows the word types that appear as the 1st words in the bigrams. i.e.

 stir<>
 soup<>
 hot<>
 salt<>
 taste<>
 add<>
 lemon<>

Thus, the word matrix can be seen as

         stir<>  soup<>  hot<>  salt<>  taste<>  add<>  lemon<>
 soup<>   5       0       12     0      0        0      0
 plate<>  0       8        9     0      0        0      0
 pepper<> 0       0        0    42      0        0      3
 salt<>   0       0        0     0     18       12      0
 lemon<>  2       0        0     0      0        3      0
 juice<>  2       0        0     0      0        0      2

When --dense is selected, word vectors displayed on stdout will look as -

  6 7
  5  0 12  0  0  0  0
  0  8  9  0  0  0  0
  0  0  0 42  0  0  3
  0  0  0  0 18 12  0
  2  0  0  0  0  3  0
  2  0  0  0  0  0  2

while by default, output will be sparse as shown by -

 6 7 12
 1  5 3 12
 2  8 3  9
 4 42 7  3
 5 18 6 12
 1  2 6  3
 1  2 7  2

If the feature file is provided, vectors are created for the given feature words only and dimensions show the words that precede them.

3. --wordorder = nocare

When wordorder is nocare, given WORD_PAIRS are treated as co-occurrence pairs and the order of words is ignored.

case (1) Feature file provided via '--feats FEATFILE' option doesnt exist.

In this case, feature and dimension files will be same and will show all unique word types encountered in the WORD_PAIRS file irrespective of the positions of the words. Each word type in WORD_PAIRS is assigned a unique index and represents the row and column of the output word matrix at that index. Thus, the output word co-occurrence matrix is square and symmetric.

Feature and dimension files for above example will show =>

 stir<>
 soup<>
 plate<>
 hot<>
 salt<>
 pepper<>
 taste<>
 add<>
 lemon<>
 juice<>

while the word matrix can be seen as

       stir<>  soup<>  plate<>  hot<>  salt<>  pepper<>  taste<>  add<> lemon<> juice<>
 stir<>   0     5       0        0      0       0         0        0      2       2
 soup<>   5     0       8       12      0       0         0        0      0       0
 plate<>  0     8       0        9      0       0         0        0      0       0
 hot<>    0    12       9        0      0       0         0        0      0       0
 salt<>   0    0        0        0      0      42        18       12      0       0
 pepper<> 0    0        0        0     42       0         0        0      3       0
 taste<>  0    0        0        0     18       0         0        0      0       0
 add<>    0    0        0        0     12       0         0        0      3       0
 lemon<>  2    0        0        0      0       3         0        3      0       2
 juice<>  2    0        0        0      0       0         0        0      2       0

Output word matrix shown on stdout will look as =>

 10 10 24
 2  5 9  2 10  2
 1  5 3  8 4 12
 2  8 4  9
 2 12 3  9
 6 42 7 18 8 12
 5 42 9  3
 5 18
 5 12 9  3
 1  2 6  3 8  3 10  2
 1  2 9  2

Or as

  10 10
  0  5  0  0  0  0  0  0  2  2
  5  0  8 12  0  0  0  0  0  0
  0  8  0  9  0  0  0  0  0  0
  0 12  9  0  0  0  0  0  0  0
  0  0  0  0  0 42 18 12  0  0
  0  0  0  0 42  0  0  0  3  0
  0  0  0  0 18  0  0  0  0  0
  0  0  0  0 12  0  0  0  3  0
  2  0  0  0  0  3  0  3  0  2
  2  0  0  0  0  0  0  0  2  0

when --dense is ON.

case (2) Feature file provided via '--feats FEATFILE' exists and lists the feature words for which the vectors are to be created.

In this case, the feature and dimension files won't be same, neither the output matrix will be square and symmetric, unless the FEATFILE is exactly same like the one automatically created by wordvec as in case (1) above. For each bigram word1<>word2 that is encountered in the WORD_PAIRS file, we check if word1 is listed in the given FEATFILE. If so, word2 is assigned a unique dimension index say j and the score of the bigram word1<>word2 is assigned to the matrix entry at [i][j], if word1 occurs at the ith position in the given FEATFILE. Then, we check if word2 is listed in the given FEATFILE and if it is and appears at the kth position in the FEATFILE, we assign a unique dimension (column) index say l to word1 and set the matrix entry at [k][l] to the co-occurrence score of the pair word1<>word2.

For example, if the FEATFILE contains -

 soup<>
 hot<>
 salt<>
 lemon<>
 pepper<>

then, the word matrix with --wordorder = nocare can be viewed as -

           stir  plate  soup  hot  pepper  salt  taste  add  juice  lemon
   soup<>    5       8     0   12       0     0      0    0      0      0
   hot<>     0       9    12    0       0     0      0    0      0      0
   salt<>    0       0     0    0      42     0     18   12      0      0
   lemon<>   2       0     0    0       3     0      0    3      2      0
   pepper<>  0       0     0    0       0    42      0    0      0      3

Output will display only the word matrix as

   5 10
   5   8   0  12   0   0   0   0   0   0
   0   9  12   0   0   0   0   0   0   0
   0   0   0   0  42   0  18  12   0   0
   2   0   0   0   3   0   0   3   2   0
   0   0   0   0   0  42   0   0   0   3

with --dense ON

and

 5 10 14
 1 5 2 8 4 12
 2 9 3 12
 5 42 7 18 8 12
 1 2 5 3 8 3 9 2
 6 42 10 3

without --dense

The dimension file created with --dims will show -

 stir<>
 plate<>
 soup<>
 hot<>
 pepper<>
 salt<>
 taste<>
 add<>
 juice<>
 lemon<>

SYSTEM REQUIREMENTS ^

AUTHORS ^

Amruta Purandare, University of Pittsburgh.

Ted Pedersen, University of Minnesota, Duluth tpederse at d.umn.edu

COPYRIGHT ^

Copyright (c) 2002-2008, Amruta Purandare and Ted Pedersen

This program is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2 of the License, or (at your option) any later version.

This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.

You should have received a copy of the GNU General Public License along with this program; if not, write to

 The Free Software Foundation, Inc.,
 59 Temple Place - Suite 330,
 Boston, MA  02111-1307, USA.
syntax highlighting: