TextClassification

This repository contains code for testing training and evaluating models on text classification data sets. Specifically, the code is currently set up to run on the following data sets:

• Reuters-21578 R52 - From http://www.daviddlewis.com/resources/testcollections/reuters21578/ with the R52 subset described at http://www.csmining.org/index.php/r52-and-r8-of-reuters-21578.html. We use the R52 subset in current experiments because we only have models that assign one label per document.

• 20 newsgroups - From http://qwone.com/~jason/20Newsgroups/.

The bulk of the work for training and evaluating the models is done using the ARKWater library (https://github.com/forkunited/ARKWater), and the code in this project is mainly for deserializing and manipulating the data-sets and loading them into the ARKWater objects.

All of the code relies on having tokenized and JSON serialized versions of the data sets. You can generate the JSON serialized data sets by downloading the original data from the URLs given above, and then running textclass.scratch.ConstructTextClassDocuments20NewsGroups or textclass.scratch.ConstructTextClassDocumentsReuters21578 on them.

Layout of the project

The code is organized into the following packages in the src directory:

• textclass.data - Classes cleaning the text classification data and performing other miscellaneous tasks.

• textclass.data.annotation - Classes for loading the text classification documents into memory.

• textclass.scratch - Code for performing miscellaneous tasks. The files in this directory contain the main functions where the code starts running.

• textclass.util - Miscellaneous utility classes.

textclass.scratch contains the entry points for the code. So if you're trying to understand how this library uses ARKWater to train and test the text classification models, you should start by looking at classes in textclass.scratch.

The experiments directory contains experiment configuration files for running experiments through the Experiment classes in textclass.scratch using ARKWater.

How to run things

Before running anything, you need to configure the project for your local setup. To configure, do the following:

1. Untar the jars at files/jars.tgz into an appropriate location.

2. Copy files/build.xml and files/textclass.properties to the top-level directory of the project.

3. Fill out the copied textclass.properties and build.xml files with the appropriate settings by replacing text surrounded by square brackets.