NAME
FAQ Frequently Asked Questions about the Ngram Statistics Package
SYNOPSIS
Frequently Asked Questions about the Ngram Statistics Package.
DESCRIPTION
This FAQ is very much a work in progress, and is written somewhat
informally. Please take the information contained herein in that light.
I'd be happy to elaborate on any point raised here that isn't clear.
Also, if you spot errors or have suggestions I'd be happy to hear about
those!
Please send your questions to me (tpederse@d.umn.edu) or the NSP mailing
list <http://groups.yahoo.com/group/ngram/>.
How should I cite NSP in a paper?
This is our preferred reference:
@inproceedings{BanerjeeP03,
author = {Banerjee, S. and Pedersen, T.},
title = {The Design, Implementation, and Use of the
{N}gram {S}tatistic {P}ackage},
booktitle = {Proceedings of the Fourth International
Conference on Intelligent Text Processing and
Computational Linguistics},
year = {2003},
month ={February},
pages = {370--381},
address = {Mexico City}}
You can get a copy of this paper at:
L<http://www.d.umn.edu/~tpederse/Pubs/cicling2003-2.pdf>
If you would also like to cite a URL, the official home page of NSP is :
L<http://ngram.sourceforge.net>
Where can I find other papers by users of NSP?
We have established an NSP bibliography that includes some papers by
users. Please make contributions to this bibliography when you publish a
paper using NSP!
L<http://www.d.umn.edu/~tpederse/nsp-bib/>
General Questions
What is the class structure of your Measure hierarchy?
The following diagram tries to lay that all out:
<http://cpansearch.perl.org/src/TPEDERSE/Text-NSP-1.13/doc/NSP-Class-dia
gram.pdf>
In general we've tried to group the measures into families that reflect
their underlying relationships, and allow us to avoid replicating a lot
of code.
Why do you have 'make test' and then csh ./ALL-TESTS.sh
This reflects the historical development of NSP. Initially NSP was based
on command line scripts, and was not easy to test using the standard
Perl methods you see in 'make test'. However, once NSP went object
oriented, using 'make test' became much easier.
We use ./ALL-TESTS.sh to test our command line programs very
extensively, so you should still run this and take note of any errors.
Please make sure you run 'csh ./ALL-TESTS.sh' AFTER running 'make
install', otherwise some of the tests cases won't find the modules they
need (since they haven't been installed and since ./ALL-TESTS.sh does
not look into the build libraries prior to installation.
Why do you use the C shell (csh) so much?
This reflects the historical development of NSP. Initially NSP was
developed on Solaris systems, which use tcsh as their default shell.
Since we do not do anything too fancy with the shell, it seemed like a
reasonable thing to stay with csh even after our site and many others
moved to Linux, where the bash shell is more common.
What is in the future for NSP?
Many things! See TODO.pod for a fairly complete accounting of our future
plans. We are always open to suggestions, so please send those to the
mailing list for consideration and discussion.
Why is NSP so slow on large files of text?
As of version 0.61 (and before) NSP does all counting using a Perl hash.
In other words, each Ngram is an element in a hash, and a counter is
kept for that Ngram. This works very nicely, but if you start to deal
with millions of words of text it can result in a very large hash that
consumes quite a bit of memory.
Remember, the crucial constraint is not how many words in the corpus,
but how many unique Ngrams you have in the text. As N gets larger, the
number of hash elements grows larger and larger. However, if you are
dealing with unigrams NSP can process extremely large files (50 million
words) quite efficiently.
To start with I would recommend that you gradually increase the value of
N and the amount of text you try to process with NSP, just to get a
sense of how long things take on your system. Start with processing
unigrams in a 1,000,000 word corpus, then try bigrams. Then move up to
5,000,000 words with unigrams, and so forth.
Questions about Significance Testing
What is Fisher's exact test?
Fisher's exact test is a significance test that is considered to be more
appropriate for sparse and skewed samples of data than statistics such
as the log likelihood ratio (G^2) or Pearson's Chi-Squared test (X^2).
The paper "Fishing for Exactness" gives a fairly detailed description of
Fisher's exact test relative to these other tests. Find it with the 1996
entries at: <http://www.d.umn.edu/~tpederse/pubs.html> It is also
available at: <http://xxx.lanl.gov/abs/cmp-lg/9608010>
What is a left sided, right sided, and two sided Fisher's exact test?
Fisher's exact test is computed by fixing the marginal totals of a 2x2
table and then determining the probability of each of the possible
tables that could result in those marginal totals. The different sided
tests vary in how these individual probabilities are summed into the
final value produced by the test.
The left sided test is recommended (see 3) since it results in a value
that is easy to interpret as a measure of association between words.
Rather loosely speaking, it is the probability of randomly sampling a
2x2 table where a bigram occurs less frequently than was observed in the
corpus you are working with. So, a high left sided probability indicates
that you are very unlikely to observe the bigram more frequently than
you already have (if you took a random sample from a similar
population). This suggests that the bigram is a "special" pair of words
that may merit further attention.
The paper "Fishing for Exactness" gives a fairly detailed description of
computing left, right and two sided values. Find it with the 1996
entries at: <http://www.d.umn.edu/~tpederse/pubs.html> It is also
available at: <http://xxx.lanl.gov/abs/cmp-lg/9608010>
Why do you recommend the use of a left sided Fisher's exact test?
The main advantage of the left sided test it is interpreted much like
the other tests that we provide. The larger the value the stronger the
association between the words in the bigram. This is the same property
as mutual information (tmi.pm), the Dice Coefficient (dice.pm),
Loglikelihood ratio (ll.pm), and Pearson's test (x2.pm). Rankings that
are done by these measures can all be compared directly, where rank 1 is
assigned to the most strongly associated bigram/s.
A right sided test is symmetric with the left sided test, so larger
values indicate weaker association between the words. There is nothing
wrong with this approach, except that you can not directly compare the
rankings of a right sided test with, for example, mutual information.
A two sided test doesn't seem to have a very natural mapping to
explaining bigram data.
Why don't the log likelihood tests (ll.pm) and Pearson's test (x2.pm) report the significance values associated with their raw scores?
We don't assign significance values to these tests since setting the
p-values seems to be a fairly arbitrary choice. It isn't clear that p
should be .01, .05, or something else. This strikes us as an important
limitation of significance tests. Rather than setting an arbitrary
cutoff on these values, we recommend looking at the raw scores from
these tests in order to establish cutoffs.
Having said all this, it is very likely that future versions of NSP will
have this capability.
How do I decide which of the tests your provide is the one I should use?
The tests provided here can be divided into three classes:
1) Power Divergence Family:
Log Likelihood ratio (ll.pm)
Pearson's Chi Squared test (x2.pm)
2) Exact test of statistical significance
Fisher's exact test (leftFisher.pm and rightFisher.pm)
3) Information theoretic measures
Mutual Information (tmi.pm)
Dice Coefficient (dice.pm)
If you would like to use a significance test with a predefined p-value
then Fisher's exact test is your choice. We have tried to implement this
as efficiently as possible, and it seems to run fairly quickly even with
large sample sizes.
If you would prefer to have a score that provides you with a measure of
the divergence of the observed from the expected values (given an
assumption of independence between the words in the bigram) then the
power divergence family is a fine choice. Note that Ll.pm and X2.pm
should provide the same results if the data is such that no asymptotic
assumptions are being violated. If you observe difference values for the
same bigram from these tests, then one of them is flawed! You can't be
sure which one it is either. There is some very fine background material
on the issue of the likelihood ratio (G^2) versus Pearson's test (X^2)
in the following:
@book{ReadC88,
author={Read, T. and Cressie, N.},
title={Goodness of fit Statistics for Discrete Multivariate Data},
year = {1988},
address = {New York, NY},
publisher = {Springer-Verlag}}
If you are interested in Mutual information, then you have it, and you
have the closely related alternative measure the Dice Coefficient.
How do I interpret the values of these test scores?
Carefully. Subjectively. Creatively.
For all of the tests, a higher score means a stronger measure of
association among the words in the bigram.
In general, we prefer to make comparisons among the tests based on the
different rankings they assign rather than the absolute scores they
return. However, you can directly compare the scores from the likelihood
ratio (Ll.pm) and Pearson's test (X2.pm). Otherwise, you should not
think of comparing these scores because they are apples and oranges. The
rank program (rank.pl) ranks the bigrams as scored by any two given
measures and will produce a table and rank correlation coefficient
showing how they compare. Please note however that you should not
include a right sided or two sided Fisher's test in such comparisons
since they do not follow the convention that a higher score means that
there is greater association between the bigrams. These tests should not
be compared to others with the rank.pl program!
Fisher's left sided test seems to rank a lot of bigrams first! Why?
As sample sizes get larger, the hypergeometric probabilities associated
with each possible 2x2 table of bigram data (given fixed marginal
totals) tend to approach 1. Consider setting the precision of the test
relatively high (10 or 15 digits) in order to observe this. When the
default setting is used (4 digits) there tends to be quite a lot of
rounding to 1.0000.
The paper "Fishing for Exactness" shows how these hypergeometric
probabilities are computed. Find it with the 1996 entries at:
<http://www.d.umn.edu/~tpederse/pubs.html> It is also available at:
<http://xxx.lanl.gov/abs/cmp-lg/9608010>
Questions about stop lists and removing non-tokens
How does the stop list option (--stop) work?
[Revised for version 0.53]
A stop list is a list of words that you don't want count.pl to count.
These words (or the Ngrams that they form) are not counted and do not
figure into the overall sample size.
Rather than specifying a list of words in a stoplist, as of version 0.53
you can specify Perl regular expressions.
There are two modes in which stop lists can be used. In the "AND" mode,
every word that makes up an N-gram must be found in the stoplist for
that N-gram to be eliminated. In the "OR" mode, any N-gram that includes
at least one word from the stoplist will be eliminated.
Your stop list should be a plain text file that has one Perl regular
expression per line.
For example, suppose your stop list consisted of:
@stop.mode=OR
/\bthe\b/
/\band\b/
/\bof\b/
[Note that \b indicates a word boundary in a Perl regex.]
Any N-gram that contains 'the', 'and', or 'of' would be excluded.
A note on counting. Suppose that the bigram "of the" occurred 700 times
in a text and was excluded by an AND mode stop list. Let's suppose for
this example that it is the only bigram excluded. The total sample size
reported by count.pl would be 700 less than without the stop list. The
frequency count for "of" occurring as the first component of a bigram
would be reduced by 700, as would the frequency count for "the"
occurring as the second component of a bigram. Note that the counts for
"of" as the second component and "the" as the first component will not
be affected by the stoplist. Note that without the stop list it will
typically be the case that the first and second position counts for a
word will be the same.
Can I have modifiers in my stoplist regular expressions?
No. :)
The documentation suggests that given a list of regular expressions such
as:
/\bis\b/
/\bthe\b/
/\ban\b/
NSP actually checks each regular expression one by one. Unfortunately if
the list of regex's is very long, this becomes too slow computationally,
and so instead we actually concatenate all the regular expressions to
form one big regex, which is then used to do the matching. For example
given the regexes above, they will be combined into a single regex, like
so:
/(\bis\b)|(\bthe\b)|(\ban\b)/
and then this regex is used to do the matching. Observe of course that
this produces exactly the same effect as if we had done the comparisons
one after another, and runs much faster. However the price we pay is
that you can't use modifiers (things outside the / /'s) while defining
the regex, since then we wouldn't be able to concatenate them together
like we are doing now.
Why is there a --stop and a --nontoken option for count.pl?
The stop option allows you to specify a stop list that eliminates Ngrams
if they are completely made up of stop words (AND mode) or if one of the
words in the Ngram is a stop word (OR mode). The effect of the stop
option is to remove Ngrams from the sample.
The nontoken option allows you to eliminate words from the text prior to
the formation of Ngrams. This processing occurs well before the stop
option, which is carried out after Ngrams have been formed.
How can I disregard n-grams that cross sentence boundaries or other punctuation marks?
For example, in :
I am here today. Where are you?
I do not want to consider "today Where are" as a 3-gram.
There is a built-in option (--newline) to disregard Ngrams across the
newline (\n), but there is not one to do the same across punctuation
marks at this time. But you can use the existing functionality to
accomplish the same!
Replace punctuation marks with new line characters and then use the
option --newLine. This will disregard Ngrams that cross newline
boundaries and thereby (in this case) cross punctuation marks.
For example, assume file "sentence.txt" is as follows:
this is a sentence. this is a sentence.
Then the following command-line script:
perl -e "while(<>){s/[.]/\n/g; print}" sentence.txt > sentence.tmp
would create the following in file "sentence.tmp":
this is a sentence
this is a sentence
(To use more punctuation marks, you'd want to put them inside the
[square brackets])
And then if you run the following:
count.pl --token token_file.txt --newLine sentence.cnt sentence.tmp
the following is produced in file "sentence.cnt":
6
a<>sentence<>2 2 2
this<>is<>2 2 2
is<>a<>2 2 2
Of course, there'd be other ways of replacing punctuation marks with
newlines. Programs like sed would probably do the trick too...
There's one problem in this approach though. Assume the following is
your original "sentence.txt" file:
this is a sentence. this is
a sentence.
Replacing '.' with \n would produce the following:
this is a sentence
this is
a sentence
And then if you use --newLine, you'd not be able to catch the second
"is<>a" bigram.
To avoid this I would suggest the following:
1. First replace all new lines with spaces
2. Then replace punctuations with new lines.
The following command-line script seems to work:
perl -e "while(<>){chomp; s/[.]/\n/g; print}" sentence.txt > sentence.tmp
Be warned though that very long lines can lead to very slow processing.
Questions about rank.pl
In rank.pl you use Spearman's rank correlation coefficient. Why not use Kendall's tau or some other correlation measure?
The reason we use Spearman's measure is that it is geared towards
correlation among ranked items. Kendall's tau can be used with the
actual values of the ranked items. However, when comparing different
kinds of tests of association such direct comparisons may not be
meaningful. For example, the Dice Coefficient and the Log-likelihood
ratio produce values that are on different scales, and you can't compare
them directly. This is why we focus on comparing ranked values, which is
why we use Spearman's.
However, when your data has numerous ties in the ranks, then you may
want to employ a method of other than Spearman's as provided via
rank.pl, which reranks the tied items but continues to use the standard
Spearman's formula. See the TODO list for more discussion on this point.
Questions about kocos.pl
What is a kth order co-occurrence?
Two words are 2nd order co-occurrences if they occur with words that
occur with a particular word. For example, suppose that we observe the
following bigrams in a corpus of text:
telephone line
busy line
telephone operator
These are first order co-occurrences.
"telephone" and "busy" are said to be second order co-occurrences of
each other since they both occur with "line". (Both are first order
co-occurrences of "line"). Then "operator" is said to be a second order
co-occurrence of "line" since it occurs with "telephone" which is a
first order co-occurrence of "line".
These relationships can be visualized as graphs or chains.
line -> telephone -> operator
kocos.pl finds such kth order relationships among words.
AUTHORS
Ted Pedersen, University of Minnesota, Duluth
tpederse at d.umn.edu
Satanjeev Banerjee, Carnegie-Mellon University
BUGS
SEE ALSO
NSP home : L<http://ngram.sourceforge.net>
NSP mailing list : L<http://groups.yahoo.com/group/ngram/>
COPYRIGHT
Copyright (C) 2000-2010 Ted Pedersen
Permission is granted to copy, distribute and/or modify this document
under the terms of the GNU Free Documentation License, Version 1.2 or
any later version published by the Free Software Foundation; with no
Invariant Sections, no Front-Cover Texts, and no Back-Cover Texts.
Note: a copy of the GNU Free Documentation License is available on the
web at <http://www.gnu.org/copyleft/fdl.html> and is included in this
distribution as FDL.txt.