Bridget McInnes >
Text-NSP-1.21 >
rank.pl

rank.pl - Calculate Spearman's Correlation on two ranked lists output by count.pl or statistic.pl

Program to calculate the rank correlation coefficient between the rankings generated by two different statistical measures on the same bigram-frequency (as output by count.pl).

This is a program that is meant to be used to compare two different statistical measures of association. Given the same set of n-grams ranked in two different ways by two different statistical measures, this program computes Spearman's rank correlation coefficient between the two rankings.

Assume we have n n-grams ranked in two ways such that each ngram has two ranks created by the two measures. Let Di represent the difference between the two ranks for ngram i. Then Spearman's rank correlation coefficient r is given by the following formula:

i=n 6 X SUM(Di^2) i=1 r = 1 - ------------ n(n^2 - 1)

That is, find the sum of the squares of the differences Di between ranks for each ngram i, and then multiply by 6 and divide by n(n^2-1). Note that when for every ngram, the two ranks of are always the same, that is the two rankings perfectly match, then Di is always 0, and so r = 1. It can be shown that when two rankings are exactly opposite to each other, that is every ngram has ranks a and n-a in the two rankings, then r = -1.

Assume that test.cnt is a list of n-grams with their frequencies as output by program count.pl. Assume that we wish to test the dis/similarity of the statistical measures 'dice' and 'x2' with respect to the n-grams contained in test.cnt. To do so, we must first rank the n-grams using these two statistical measures using program statistic.pl.

perl statistic.pl dice test.dice test.cnt perl statistic.pl x2 test.x2 test.cnt

Having obtained two different rankings of the n-grams in test.cnt in files test.dice and test.x2, we can now compute the Spearman's rank correlation coefficient using these two rankings like so:

perl rank.pl test.dice test.x2.

This will output a floating point number between -1 and 1. A return of '1' indicates a perfect match in rankings, '-1' a completely reversed ranking and '0' a pair of rankings that are completely unrelated to each other. Numbers that lie between these numbers indicate various degrees of relatedness / un-relatedness / reverse-relatedness.

Recall that program statistic.pl ranks n-grams in such a way that the fact that an ngram has a rank 'r' implies that there are 'r-1' distinct scores greater than the score of this ngram. Thus say if 'k' n-grams are tied at a score with rank 'a', then the next highest scoring n-grams is given a rank 'a+1' instead of 'a+k+1'.

For example, observe the following file output by statistic.pl:

11 of<>text<>1 1.0000 2 2 2 and<>a<>1 1.0000 1 1 1 a<>third<>1 1.0000 1 1 1 text<>second<>1 1.0000 1 1 1 line<>of<>2 0.8000 2 3 2 third<>line<>3 0.5000 1 1 3 line<>and<>3 0.5000 1 3 1 second<>line<>3 0.5000 1 1 3 first<>line<>3 0.5000 1 1 3

Observe that although 4 bigrams have a rank of 1, the next highest scoring bigram is not ranked 5, but instead 2.

Spearman's rank correlation coefficient requires the more conventional kind of ranking. Thus the above file is first "re-ranked" to the following:

11 of<>text<>1 1.0000 2 2 2 and<>a<>1 1.0000 1 1 1 a<>third<>1 1.0000 1 1 1 text<>second<>1 1.0000 1 1 1 line<>of<>5 0.8000 2 3 2 third<>line<>6 0.5000 1 1 3 line<>and<>6 0.5000 1 3 1 second<>line<>6 0.5000 1 1 3 first<>line<>6 0.5000 1 1 3

And then these rankings are used to compute the correlation coefficient.

The two input files to rank.pl may not have the same set of n-grams. In particular, if one or both of the files generated using statistic.pl has been generated using a frequency, rank or score cut-off, then it is likely that the two files will have different sets of n-grams. In such a situation, n-grams that do not occur in both files are removed, the n-grams that remain are re-ranked and then the correlation coefficient is computed.

For example assume the following two files output by statistic.pl using two fictitious statistical measures from a fictitious file output by program count.pl.

The first file:

first<>bigram<>1 4.000 1 1 second<>bigram<>2 3.000 2 2 extra<>bigram1<>3 2.000 3 3 third<>bigram<>4 1.000 4 4

The second file:

second<>bigram<>1 4.000 2 2 extra<>bigram2<>2 3.000 4 4 first<>bigram<>3 2.000 1 1 third<>bigram<>4 1.000 3 3

Observe that the bigrams extra<>bigram1<> in the first file and extra<>bigram2<> in the second file are not present in both files. After removing these bigrams and re-ranking the rest, we get the following files:

The modified first file:

first<>bigram<>1 4.000 1 1 second<>bigram<>2 3.000 2 2 third<>bigram<>3 1.000 4 4

The modified second file:

second<>bigram<>1 4.000 2 2 first<>bigram<>2 2.000 1 1 third<>bigram<>3 1.000 3 3

Since each ngram belongs to both files, the correlation coefficient may be computed on both files.

We provide c-shell script rank-script.sh that takes a bigram count file and the names of two libraries and then computes the Spearman's rank correlation coefficient by making use successively of programs statistic.pl and rank.pl.

Run this script like so: rank-script.sh <lib1> <lib2> <file>

where <lib1> is the first library, say dice <lib2> is the second library, say x2 <file> is the file of ngrams and their frequencies produced by program count.pl.

For example, if test.cnt contains bigrams and their frequencies, we can run it like so to compute the rank correlation coefficient between dice and x2:

csh rank-script.sh dice x2 test.cnt.

This runs the following commands in succession:

perl statistic.pl dice out1 test.cnt perl statistic.pl x2 out2 test.cnt perl rank.pl out1 out2

The intermediate files out1 and out2 are later destroyed.

Note that since no command line options are utilized in the running of program statistic.pl here, this script only works for bigrams and enforces no cut-offs. However the script is simple enough to be manually modified to the user's requirements.

Ted Pedersen, tpederse@umn.edu Satanjeev Banerjee, bane0025@d.umn.edu

This work has been partially supported by a National Science Foundation Faculty Early CAREER Development award (\#0092784) and by a Grant-in-Aid of Research, Artistry and Scholarship from the Office of the Vice President for Research and the Dean of the Graduate School of the University of Minnesota.

Copyright (C) 2000-2003, Ted Pedersen and Satanjeev Banerjee

This suite of programs 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.

Note: The text of the GNU General Public License is provided in the file GPL.txt that you should have received with this distribution.

syntax highlighting: