Name

README.Demos.pod

Description

The Demos directory allows a user to run various sample experiments with the SenseClusters system. Sample data is provided in the /Data directory, and there are scripts available that show how to exercise some of the major functionality of the package.

The package supports two basic models, native SenseClusters and Latent Semantic Analysis. Both of those are demoed via scripts found in this directory.

SenseClusters provides a wrapper program called discriminate.pl that can be used to run many different experiments. This wrapper calls many of the other programs found in the package and integrates their functionality for you. We show how discriminate.pl can be used to carry out target word discrimination both in the native SenseClusters mode as well as for Latent Semantic Analysis (target-wrapper.sh).

SenseClusters also allows a user to customize the sequence of operations carried out in their experiments, and examples of that are shown in scripts for native SenseClusters (sc-toolkit.sh) and Latent Semantic Analysis (coming soon).

SenseClusters also supports word clustering in native SenseClusters mode and feature clustering in LSA mode. These are shown in the script word-wrapper.sh

Data preparation

A number of sample data files are provided in /Data. The files that begin with eng-lex-sample.* are from the Senseval-2 word sense disambiguation exercise, and consist of multiple contexts that contain a given target word that will be clustered.

The training data file (eng-lex-sample.training.xml) is intended to be used for feature selection. The evaluation/test file (eng-lex-sample.evaluation.xml) is the data that will be clustered. eng-lex-sample.key is the official key from the Senseval-2 event, which indicates what the correct sense assignment of each context should be (according to that event at least). This information can be used for evaluation. Finally, the file eng-global-train.txt is a file of raw text that can be used for feature selection in the global mode, that is without respect to any particular target words.

Before you start working with the data, you should run the following script (makedata.sh):

makedata.sh

Creates experimental data in dir LexSample.

makedata also takes advantage of a SenseClusters preprocessing wrapper program setup.pl, and allows filtering of test data based on the frequency of the senses found in that data.

Note

Before running any demo, make sure to

remove LexSample directory if one already exists (rm -fr LexSample)

run makedata.sh to create fresh copy of LexSample

Acknowledgments

This work has been partially supported by a National Science Foundation Faculty Early CAREER Development award (#0092784).