NAME
SenseClusters
SYNOPSIS
SenseClusters is a suite of Perl programs that supports unsupervised clustering of similar contexts. It relies on it's own native methodology, and also provides support for Latent Semantic Analysis.
SenseClusters is a complete system that takes users from preprocessing of raw text to providing clustered output. It supports the selection of features, the creation of various kinds of context representations, dimensionality reduction by singular value decomposition, clustering, and analysis of results.
SenseClusters integrates specialized tools such as the Ngram Statistics Package (NSP), SVDPACK, the Perl Data Language (PDL) and CLUTO to provide a variety of choices and high efficiency at each step in its processing.
OVERVIEW
SenseClusters supports several different methods of clustering contexts. These include the native SenseClusters methodology, which is based on the use of first and second order representations of contexts. It also includes support for clustering lexical features using the native SenseClusters methodology or Latent Semantic Analysis.
SenseClusters is based strictly on lexical features and does not rely on any manually created training data or external knowledge sources, and as such is language independent. The only requirement is that the language should be able to be tokenized via Perl regular expressions, which can be specified by the user. In fact, tokenization is so flexible that features could consist of characters, pairs of characters, etc.
SenseClusters can be applied to the problem of discriminating word meanings or ambiguous names, using the target or head word representation. This is sometimes also called "headed" data, where each context is centered around the given target whose meanings are to be discovered. In this case the contexts that contain the given target word are clustered, and each cluster is assumed to correspond to a different meaning of that word.
SenseClusters can also be applied to the problem of grouping short units of text that have no target or head (which is sometimes referred to as a "headless" representation. In this case there is no head or center to the context, so the entire context is being clustered to determine the meaning or topic of the context as a whole. Email categorization or news article clustering are examples of problems that could be approached using headless data.
SenseClusters will automatically determine the number of clusters in the data based on a number of different automatic stopping measures we have developed, three of which are based on clustering criterion function, and one which is an adaptation of the well-known Gap Statistic.
SenseClusters can also be applied to the problem of clustering words or lexical features, in hopes of discovering synonyms, antonyms, or other classes of words.
Broadly speaking, SenseClusters can be used for any task that requires the recognition of contextually similar units of text, or words that occur in similar contexts.
DOCUMENTATION
SenseClusters' documentation is available ONLINE at : http://senseclusters.sourceforge.net/SenseClusters-Code-README.html
For OFFLINE browsing, directory Docs/HTML is provided in SenseClusters' main package directory and the SenseClusters-Code-README.html file can be found here and locally browsed.
All programs have inline source code documentation written in pod style and this can be browsed from command line as a man page or using the 'perldoc' command. For example, 'man bitsimat.pl' or 'perldoc bitsimat.pl' will displayed the documentation for the bitsimat.pl program. Each program also has a --help option to provide information about program options.
GETTING STARTED
You might first like to run the Demo scripts in Demos/ directory to get an idea of SenseClusters' usage and functionality, or try the web interface that is provided at http://senseclusters.sourceforge.net.
Demos/ contains scripts that utilize the wrapper program discriminate.pl that calls various other programs from the package to run a complete experiment. It also contains examples where specialized experiments are constructed directly from the programs provided in the package. In general it would be useful to consult the flowcharts in Docs/Flows to understand the overall structure of the package.
The web interface provides an intuitive means of formulating and running discriminate.pl commands, so the use of the web interface and certainly be instructive in terms of how to formulate discriminate.pl commands.
The contexts that you wish to cluster must be in Senseval-2 format. This is a simple XML markup that indicates the beginning and end of each context, and allows you to specify a target word and a "correct" categorization of the context, if you know that information. There is a pre-processing program text2sval.pl in Toolkit/preprocess/plain/ that converts plain text data (with a single context on each line) into Senseval-2 format. There is also a large amount of sample data that is already in Senseval-2 format available at http://senseclusters.sourceforge.net
You can also (optionally) provide a separate training file in plain text format to be used as the feature selection data. If you don't do this, then the features will be selected from the contexts to be clustered.
PACKAGE ORGANIZATION
After downloading and unpacking SenseClusters, you should find following files/directories within SenseClusters' directory.
README.SC.pod
This file.
INSTALL
The installation guide, which lists all package dependencies.
discriminate.pl
A wrapper program that acts as a driver for many other programs in the package. It clusters the given text instances based on their contextual similarities.
Demos/
A directory of scripts that demonstrate SenseClusters' usage and functionality.
Toolkit/
A directory of Perl programs implemented and used by SenseClusters. Users who are interested to use SenseClusters' tools individually and separately without using the wrapper programs are encouraged to browse through the Toolkit and Toolkit.pod.
Docs/
A directory of SenseClusters' documentation in html format.
Directory Docs/Flows/ contains flow diagrams that illustrate how to put together the programs provided in SenseClusters' Toolkit with other packages like NSP, SVDPACK and CLUTO to run experiments without wrappers.
Testing/
A directory of test cases written as C-shell scripts that will test if the package is installed properly or not.
Web/
Contains an easy to use and install web interface for SenseClusters.
Changes/
A directory of changelogs that document the changes and improvements done during each version.
Makefile.PL
Generates a Makefile on running 'perl Makefile.PL'.
GPL.txt
A copy of the GNU General Public License, the terms under which SenseClusters is distributed.
FDL.txt
A copy of the GNU Free Documentation License, the terms under which the documentation of SenseClusters is distributed.
CONTACT US
SenseClusters was originally developed and maintained by Amruta Purandare and Ted Pedersen from September 2002 until August 2004. Since that time it has been developed and maintained by Anagha Kulkarni and Ted Pedersen.
Please join our mailing lists to participate in the package related discussions, to post your questions or bugs and also to suggest enhancements to the package functionality.
To subscribe to the user's mailing list, visit : http://lists.sourceforge.net/lists/listinfo/senseclusters-users
To subscribe to a low volume news mailing list, visit : http://lists.sourceforge.net/lists/listinfo/senseclusters-news
To subscribe to the developer's mailing list, visit : http://lists.sourceforge.net/lists/listinfo/senseclusters-developers
Recent version of SenseClusters can be downloaded from : http://senseclusters.sourceforge.net/
SEE ALSO
SenseClusters' ONLINE Documentation at http://senseclusters.sourceforge.net/SenseClusters-Code-README.html
AUTHORS
Ted Pedersen
University of Minnesota, Duluth
tpederse@d.umn.edu
http://www.d.umn.edu/~tpederse/
Amruta Purandare
University of Pittsburgh
amruta@cs.pitt.edu
http://www.cs.pitt.edu/~amruta/
Anagha Kulkarni
University of Minnesota, Duluth
kulka020@d.umn.edu
http://www.d.umn.edu/~kulka020/
Mahesh Joshi
University of Minnesota, Duluth
joshi031@d.umn.edu
http://www.d.umn.edu/~joshi031/ACKNOWLEDGMENTS
This work has been partially supported by a National Science Foundation Faculty Early CAREER Development award (Grant #0092784).
We would also like to express our special thanks to :
Dr. George Karypis and his research group for developing CLUTO, Dr. Michael Berry and the co-developers of SVDPACK and SVDPACKC, Christian Soeller and the PDL developers' team, and Satanjeev Banerjee for developing the Ngram Statistics Package.
COPYRIGHT
Copyright (c) 2003-2006, Ted Pedersen, Amruta Purandare, Anagha Kulkarni, and Mahesh Joshi
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.