SenseClusters - Cluster similar contexts using co-occurrence matrices and Latent Semantic Analysis


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 (Text::NSP), SVDPACK, the Perl Data Language (PDL) and CLUTO to provide a variety of choices and high efficiency at each step in its processing.


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.


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' or 'perldoc' will displayed the documentation for the program. Each program also has a --help option to provide information about program options.

You can see all of the modules and their associated documentation at README.Toolkit.


You might first like to run the demonstration scripts in samples/ directory to get an idea of SenseClusters' usage and functionality, or try the web interface that is provided at

samples/ contains scripts that utilize the wrapper program 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 doc/Flowcharts to understand the overall structure of the package.

The web interface provides an intuitive means of formulating and running commands, so the use of the web interface and certainly be instructive in terms of how to formulate 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 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

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.


After downloading and unpacking SenseClusters, you should find following files/directories within SenseClusters' directory.


Read-only copies of documentation found in doc/*.pod


A copy of the GNU General Public License, the terms under which SenseClusters is distributed.


A copy of the GNU Free Documentation License, the terms under which the documentation of SenseClusters is distributed.

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.


Generates a Makefile on running 'perl Makefile.PL'.


Contains various *.pod files that are kept in a read only form in the top level directory.

doc/Flowcharts/ 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.


A stub for the Text::Similarity perl module. At present SenseClusters is oriented about the command line, so this is mostly for the benefit of CPAN indexing.


A stub directory created by h2xs - future site of test cases rather than /Testing


A directory of scripts that demonstrate SenseClusters' usage and functionality.


Contains a modified version of SVDPACKC, and a script that can be run to automatically install it, and retrieve and install Cluto.


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.


Contains an easy to use and install web interface for SenseClusters.


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 :


To subscribe to a low volume news mailing list, visit :


To subscribe to the developer's mailing list, visit :


Visit the SenseClusters project home page :



The SenseClusters project has been partially supported by a National Science Foundation Faculty Early CAREER Development award (Grant #0092784). This award funded the work of Amruta Purandare (2002-2004) and Anagha Kulkarni (2004-2006).


 Ted Pedersen
 University of Minnesota, Duluth
 tpederse at

 Amruta Purandare
 University of Pittsburgh

 Anagha Kulkarni
 Carnegie-Mellon University

 Mahesh Joshi
 Carnegie-Mellon University


Copyright (c) 2003-2008, 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.

1 POD Error

The following errors were encountered while parsing the POD:

Around line 186:

Expected text after =item, not a bullet