Skip to content

An English version of the context based spell checker Valkuil.net.

Notifications You must be signed in to change notification settings

Woseseltops/fowlt

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

fowlt

An English version of the context based spell checker Valkuil.net.

DESCRIPTION

Fowlt.net is the English version of the Dutch context based spell checker Valkuil.net (available on http://www.valkuil.net). A context based spell checker tries to find and correct errors by comparing bits of a text to the large data set it was trained on. If a word is not expected on the basis of this training data, Valkuil.net marks it as an error. Because of this, Fowlt.net can make smarter decisions than standard spell checkers.

HOW TO INSTALL

  1. Download Fowlt. Fowlt can be used on any Linux pc with Python 2. To obtain Fowlt from GitHub:

$ git clone git://github.com/Woseseltops/fowlt.git

  1. Install PyNLPl. Fowlt makes use of the PyNLPl Python library, which can be installed with '$ easy_install pynlpl'. In case you don't have permission to use easy_install, you can also clone PyNLPl from GitHub into the Fowlt directory (https://github.com/proycon/pynlpl).

  2. Install ucto, timbl, timblserver and WOPR. Fowlt makes use of already existing NLP and machine learning software. Ucto (http://ilk.uvt.nl/ucto/), timbl and timblserver (http://ilk.uvt.nl/timbl/) can be installed easily. Wopr (http://ilk.uvt.nl/wopr/) has to be compiled by the user.

  3. Compile the modules. Several modules need to be compiled before use. This can be done automatically by doing a make command from the main directory.

  4. Configure the settings. You can tell the program where Timblserver and WOPR are installed by editing the file servers/server_settings. This file also has the option 'wopr_large_corpus' set to 0, which means that only a small training set will be used (10000 lines of text). If you are running Fowlt from a place with a lot of memory, you can use the large training set (1000000 lines of text) by setting this option to 1.

HOW TO USE

  1. Start the servers. For its context-based modules Fowlt uses two servers that should be running. One server is for the so-called WoprCheckerModule, the other is for all modules that check for confusibles. These servers can be started by running servers/start_woprserver.py and servers/start_timblserver.py respectively. If you don't start these servers, Fowlt will run, but without its most important functionality. If you run the servers, but not at the localhost, you should change that in the file client_settings, found in the main directory.

  2. Start the processchain. Once the servers are running, you can use Fowlt itself by running fowlt_processchain.py. This script currently takes only one argument, which is the name of the document you want to correct. This can be any document that contains plain UTF-8 text.

  3. If you're ready, stop the servers. The Woprserver stops automatically if you interrupt the program in one way or the other, but the Timblserver has to be killed actively. This can be done by running servers/stoptimblserver.py.

HOW TO UPDATE

To keep Fowlt up to date, regularly check back on GitHub for changes and obtain the latest using a simple:

fowlt$ git pull

HOW TO EXPAND

If you want, you can add your own modules to Fowlt. Each module is implemented as a class in fowlt_processchain.py. This is all very straightforward, so you can probably get all info needed by simply looking at the other examples. If you want to create a confusible checker module similar to the ones Fowlt uses, but with another confusible, you need to know a little more. It is done as follows:

  1. Make sure you have a file with lots of text to train Fowlt on. The British National Corpus was used for the modules included in the standard Fowlt installation. You don't have to exclude the utterances that aren't relevant for what you want to train the model for; Fowlt will select all relevant material automatically.

  2. Create a language model. This can be done automatically by running 'confusibletrainer/confusible_trainer.py [word1,word2,word3] [corpus]'. The first argument is a comma-separated list of words you want to train your module on, the second argument is the corpus you selected. Three files will be created: a .inst file (the training material), an IGTree file and IGTree.wgt file (the actual language model created by Timbl). If you want balanced material (the same amount of examples for each confusible option), please add '-balanced' to the command.

  3. Evaluate the model. Before adding your new module to Fowlt, you might want to know how good it is. This can be tested by giving the .inst file to confusibletester/confusible_tester.py. This script evaluates to what extent timbl will be able to predict words and detect errors on the basis of this training material. It generates a lot of files, but you will mainly interested in the .output file, which contains info on accuracy, precision, recall, etc. The other files are for evaluating purposes and should not be used in your module.

  4. Add your new module to Fowlt. You have to add your new module both to the Fowlt server and the Fowlt client. The server can be updated by adding the line 'confusible1-confusible2="-a1 +vdb +D -i timbl_servers/confusible1,confusible2.IGTree"' to servers/timblservers/confusibles.conf (and replace 'confusible1' and 'confusible2' with the actual confusibles you trained for, obviously). The client can be updated by subclassing AbstractModule in fowlt_processchain.py.

About

An English version of the context based spell checker Valkuil.net.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published