TODO Things to do in the Ngram Statistics Package
Ngram Statistics Package Todo list
The following list describes some of the features that we'd like to include in NSP in future. No particular priority is assigned to these items - they are all things we've discussed amongst ourselves or with users and agree would be good to add.
If you have additional ideas, or would like to comment on something on the current list, please let us know via the ngram mailing list.
It would be nice to offer a web interface or web server for users who just want to run a few measures.
NSP is geared for the Roman alphabet (Latin-1). Perl has increasingly better Unicode support with each passing release, and we will incorporate Unicode support in future. We attempted to use the Unicode features in Perl 5.6, but found them to be incomplete. We have not yet attempted this with Perl 5.8 (the now current version) but it is said to be considerably better.
Perl support for unicode will include language / alphabet specific definitions of regular expression character classes like \d+ or \w+ (digits and non-white space characters). So you should be able to use (in theory) the same regular expression definitions with any alphabet and have it match in a way that makes sense for that language.
Our expertise in this area is fairly limited, so please let us know if we are missing something obvious or misunderstanding what Perl is attempting to do.
In a discussion in Feb 2008, Richard Jelinek suggested the use of the ENCODE module, discussion starts here :
In that discussion some drawbacks to 'use locale' were pointed out, so for the moment we have made no changes, but it seems like fitting NSP with ENCODE support is a good idea.
Right now all the ngrams being counted are stored in memory. Each ngram is an element in a hash. This is ok for up to a few million word corpora, but after that things really slow down. We would like to pursue the idea of using suffix trees which would greatly improve space utilization.
The use of suffix trees for counting term frequencies is based on :
Yamamoto, M. and Church, K (2001) Using Suffix Arrays to compute Term Frequency and Document Frequency for All Substrings in a Corpus, Computational Linguistics, vol 27:1, pp. 1-30, MIT Press.
Find the article at:
In fact, they even provide a C implementation:
However, we would convert this into Perl and may need to modify it somewhat to fit into NSP.
Another alternative would be to simply modify the count.pl program such that rather than using memory it used disk space to accumulate counts. This would be very slow but might suffice for certain situations. This is what huge-count.pl currently does.
Another alternative would be to tie the hashes that are used in NSP to a database, and thereby reduce some memory use.
Regardless of the changes we make to counting, would continue to support counting in memory, which is perfectly adequate for smaller amounts of corpora.
The web is a huge source of text, and we could get counts for words or ngrams from the web (probably using something like Perl LWP module).
Rather than running count.pl on a particular body of text (as is the case now) we'd probably have to run count.pl such that it looked for counts for a specific set of words as found on the web. Simply running count.pl on the entire www wouldn't really make sense. So perhaps we would run count on one sample to get a list of the word types/ngrams that we are interested in, and then run count on the www to find out their respective counts.
[Our interest in this has been inspired by both Peter Turney (ACL-02 paper) and Frank Keller (EMNLP-02 paper).]
Counting words and ngrams in large corpora could be parallelized. The trick is not so much in the counting, but in the combining of counts from various sources.
This is something we might try and implement using MPI (Message Passing Interface).
When processing large files, count.pl gives no indication of how much of the file has been processed, or even if it is still making progress. A "progress meter" could show how much of the file has been proceeded, or how many ngrams have been counted, or something to indicate that progress is being made.
If count.pl encounters a very long line of text (with literally thousands and thousands of words on a single line) it may operate very very slowly. It would be good to let a user know that an overly long line (we'd need to define more precisely what "overly long" is) is being processed (this fits into the progress meter mentioned above) so that a user can decide if they want to continue with this, or possibly terminate processing and reformat the input file.
The --newLine switch tells count.pl that Ngrams may not cross over end of line markers. Presumably this would be used when each line of text consists of a sentence (thus the end of a line also marks the end of a sentence). However, if the text is not formatted and there may be multiple sentences per line, or sentences may extend across several lines, we may want to allow --newLine to include other characters that Ngrams would not be allowed to cross.
For example we could have the switch --dontCross "\n\.,;\?" which would prevent ngrams from crossing the newline, the fullstop, the comma, the semicolon and the question mark.
Our current --recurse option creates a single count output file for all the words in all the texts found in a directory structure. We might want to be able to process all the files in a directory structure such that each file is treated separately and a separate count file is created for it.
For example, suppose we have the directory /txts that contains the files text1 and text2.
count.pl --recurse output txts
output will consist of the combined counts from txts/text1 and txts/text2.
This new option would count these files separately and produce separate count output files.
DONE IN VERSION 1.13! (--uremove option): What about having a frequency cutoff for count.pl that removed any ngrams that occur more than some number of times? The idea here would be to eliminate high frequency ngrams not through the use of a stoplist but rather through a frequency cutoff, based on the presumption that most very high frequent ngrams will be made up of stop words.
What about a percentage cutoff? In other words, eliminate the least (or most) frequent ngrams?
It would be useful to allow NSP to automatically create a stoplist based on a combination of frequency counts and/or scores like tf/idf. While tf/idf depends on the idea of a document, we would simply chunk up a large corpus into 100 token long pieces, and consider each piece a document, and consider stop words those words that occur in some number of these chunks.
Right now count.pl and statistic.pl operate such that the output file is designated first, followed by the input file.
count.pl outputfile inputfile
However, there are advantages to allowing a user to redirect input and output, particularly in the Unix and MS-DOS world. As Derek Jones pointed out to us, if we have Windows users they are probably looking for a GUI (and they won't find much will they!!). This would enable the use of syntax such as...
count.pl input > out cat input | count.pl > outfile
which would help in building scripts, etc.
Rather than have user set paths, have a script that would ask the users questions to set things up properly. This might be especially useful if we want to maintain the "old" style of output input file specifications in count.pl and statistic.pl (see point above) as well as STDIN STDOUT. (Maybe a user could pick which one?) In addition, there may be other options that a user could specify this way (such as a default token definition, home directory, etc.)
At present huge-count.pl is only able to count bigrams. It would be very useful to extend it so that it could count Ngrams in general. Also, there is no support for windowing provided at present, so the bigrams it counts must be adjacent. It would be desirable to support windowing for bigrams and Ngrams generally.
At present all programs simply exit when they encounter an error. We will return an error code that can be detected by the calling program, so that abnormal termination is clear. This affects count.pl and statistic.pl particularly, but will also be changed in rank.pl, combig.pl and kocos.pl.
There is a certain amount of redundant code in count.pl, huge-count.pl and kocos.pl. It would be useful to make these more modular, to allow for inheritance and code sharing, as well as the use of objects (potentially).
Right now rank.pl does not handle ties in any way other than re-ranking them such that all members of a tie have the same rank, and that the next rank after the ties is incremented by the number of ties. Some sources advocate using Pearson's correlation coefficient on the ranks in case of ties :
Other sources prefer the use of Kendall's Tau over Spearman's:
Our suggestion is that if you have data with numerous ties, you want to look very carefully at alternatives to the methods described in rank.pl However, typical collocation data collected from corpora usually doesn't have too many ties, so in general we feel rank.pl remains useful.
USAGE.pod has not been updated since 2001, and is very basic.
Ted Pedersen, firstname.lastname@example.org
Last Updated : $Id: TODO.pod,v 1.6 2010/03/04 04:08:16 tpederse Exp $
home page: L<http://www.d.umn.edu/~tpederse/nsp.html> mailing list: L<http://groups.yahoo.com/group/ngram/>
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.