Ted Pedersen

NAME

modules - [documentation] Overview of WordNet::Similarity measure modules

DESCRIPTION

All the modules that will be installed in the Perl system directory are present in the '/lib' directory tree of the package. These include the semantic relatedness modules -- jcn.pm, res.pm, lin.pm, lch.pm, hso.pm, lesk.pm, vector_pairs.pm, wup.pm, path.pm and random.pm -- present in the /lib/WordNet/Similarity subdirectory and the supporting modules PathFinder, ICFinder, DepthFinder, GlossFinder, et_wn_info.pm, and string_compare.pm. There also exists a WordNet/Similarity.pm module that currently serves as a base class for all the relatedness modules and contains Perl documentation. All these modules, once installed in the Perl system directory, can be directly used by Perl programs.

A UML diagram showing how all the classes are organized is available from

http://www.d.umn.edu/~tpederse/Docs/wn-similarity-v0.13-uml.pdf

Using the relatedness modules

The semantic relatedness modules in this distribution are built as classes that define the following methods:

  new()
  getRelatedness()
  getError()
  getTraceString()

new()

The first thing that is done in order to use one of the semantic relatedness measures is to create an object of the measure. This is done by calling the 'new' method of that measure or module. For all the semantic relatedness measures provided in this package, the 'new' method takes two parameters --

(a)

a WordNet::QueryData object (REQUIRED)

(b)

the name of a configuration file for that module (Optional)

This method initializes an object of the requested measure, using the configuration file data, or with default values if a configuration file is not provided. A reference to this object is returned by the 'new' method and must be saved by the calling program, if any of the other methods of this module are to be called. It is possible to create multiple objects of the same module (possibly initialized differently by specifying different configuration files for each). The format of the configuration files is discussed later in this section.

An 'undef' value returned by the 'new' method, indicates that it was unable to create an object. It is also possible that non-fatal errors occur during the creation of the object. In such a case an object is created by the 'new' method using default conditions. However, a non-fatal error condition flag is set within the object, which can be retrieved using the getError() method. It is advisable to check for this error condition after the creation of every such object.

getRelatedness()

The 'getRelatedness' method is called on the created object to determine the semantic relatedness of two concepts (synsets in WordNet) as computed by that measure. The input parameters are two WordNet synsets, represented in the word#pos#sense format returned/used by WordNet::QueryData. In this format each synset is represented by a word from that synset, its part-of-speech and its sense number. For example, if the second sense of 'teacher' as a noun occurs in a synset containing synonyms for 'teacher', then this synset can be represented by the string 'teacher#n#2'. The 'getRelatedness' method takes as input two strings of this form and returns a floating point value, which is the semantic relatedness of these (as computed by the measure).

getError()

During a call to either the 'new' method or the 'getRelatedness' method of a measure, if a fatal or non-fatal error occurs, the module sets an error flag within the created object and sets an error string within (the exception to this is when the module is unable to create an object upon a call to the 'new' method, in which case it simply returns 'undef'). Both the error condition flag and the error string can be retrieved using the 'getError' method on the created object. The method is called without any parameters and it returns an array containing the error flag as the first element and the error string as the second element. The error flag can take the values 0, 1 or 2. A value of 0 indicates that there was no error or warning since the last call to 'getError'. 1 indicates that there was/were non-fatal error(s) (warnings) since the last call to 'getError'. A value of 2 usually indicates that the errors were serious enough to warrant the termination of the program. However, how these errors are handled is completely upto the programmer writing the Perl program. It is advisable that the error flag be checked after every call to either 'new' or 'getRelatedness', but this is not a necessary step and the error condition may be tested at less regular intervals also.

getTraceString()

If traces are enabled, a trace string generated during the last call to the 'getRelatedness' method is stored within the object. This trace string can be retrieved using the 'getTraceString' method. This method is called with no parameters and returns a scalar containing the most recently generated trace string. By default traces are not enabled. Traces can be enabled by specifying this as an option in the configuration file for the measure. Instructions for writing configuration files for the measures follow later in this section.

Examples of typical usage

To create an object of the Resnik measure, we would have the following lines of code in the Perl program.

   use WordNet::Similarity::res;
   $object = WordNet::Similarity::res->new($wn, '/home/sid/resnik.conf');

The reference of the initialized object is stored in the scalar variable '$object'. '$wn' contains a WordNet::QueryData object that should have been created earlier in the program. The second parameter to the 'new' method is the path of the configuration file for the resnik measure. If the 'new' method is unable to create the object, '$object' would be undefined. This, as well as any other error/warning may be tested.

   die "Unable to create resnik object.\n" if(!defined $object);
   ($err, $errString) = $object->getError();
   die $errString."\n" if($err);

To create a Leacock-Chodorow measure object, using default values, i.e. no configuration file, we would have the following:

   use WordNet::Similarity::lch;
   $measure = WordNet::Similarity::lch->new($wn);

To find the semantic relatedness of the first sense of the noun 'car' and the second sense of the noun 'bus' using the resnik measure, we would write the following piece of code:

   $relatedness = $object->getRelatedness('car#n#1', 'bus#n#2');

To get traces for the above computation:

   print $object->getTraceString();

However, traces must be enabled using configuration files. By default traces are turned off.

Configuration files

The behavior of the measures of semantic relatedness can be controlled by using configuration files. These configuration files specify how certain parameters are initialized within the object. A configuration file may be specified as a parameter during the creation of an object using the new method.

The configuration files follow a fixed file format. Every configuration file starts with the name of the module ON THE FIRST LINE of the file. For example, a configuration file for the Resnik module will have on the first line 'WordNet::Similarity::res'. This is followed by the various parameters, each on a new line and having the form 'name::value'. The 'value' of a parameter is optional (in case of boolean parameters). In case 'value' is omitted, we would have just 'name::' on that line. Comments are supported in the configuration file. Anything following a '#' is ignored in the configuration file.

Sample configuration files are present in the '/samples' subdirectory of the package. Each of the modules has specific parameters that can be set/reset using the configuration files. Please read the manpages or the perldocs of the respective modules for details on the parameters specific to each of the modules. For instance, 'man WordNet::Similarity::res' or 'perldoc WordNet::Similarity::res' should display the documentation for the Resnik module.

Information Content

Three of the measures provided within the package require information content values of concepts (WordNet synsets) for computing the semantic relatedness of concepts. Resnik (1995) describes a method for computing the information content of concepts from large corpora of text. In order to compute information content of concepts, according to the method described in the paper, we require the frequency of occurrence of every concept in a large corpus of text. We provide these frequency counts to the three measures (Resnik, Jiang-Conrath and Lin measures) in files that we call information content files. These files contain a list of WordNet synset offsets along with their part of speech and frequency count. The files are also used to determine the topmost node of the noun and verb 'is-a' hierarchies in WordNet. The information content file that should be used by a module is specified in the configuration file of that module. If no information content file is specified, then the default information content file, generated at the time of the installation of the WordNet::Similarity modules, is used. A description of the format of these files follows. The FIRST LINE of this file must contain the version of WordNet that the file was created with. This should be present as a string of the form

 wnver::<hashcode>

For example, if WordNet version 2.1 with a hash-code LL1BZMsWkr0YOuiewfbiL656+Q4 was used for creation of the information content file, the following line would be present at the start of the information content file.

 wnver::LL1BZMsWkr0YOuiewfbiL656+Q4

The rest of the file contains on each line a WordNet synset offset, part-of-speech and a frequency count, in the form

 <offset><part-of-speech> <frequency> [ROOT]

without any leading or trailing spaces. For example, one of the lines of an information content file may be as follows.

 63723n 667

where '63723' is a 'noun' synset offset and 667 is its frequency count. Suppose the noun synset with offset 1740 is the root node of one of the noun taxonomies and has a frequency count of 17625. Then this synset would appear in an information content file as follows:

 1740n 17625 ROOT

The ROOT tags are extremely significant in determining the top of the hierarchies and must not be omitted. Typically, frequency counts for the noun and verb hierarchies are present in each information content file. A number of support programs to generate these files from various corpora are present in the '/utils' directory of the package. A sample information content file has been provided in the '/samples' directory of the package.

NOTE: Using the "Resnik" counting it is possible to get fractional values for the frequency counts.

Depths Files

The lch and wup measures make use of auxiliary files to determine certain types of depths information. The lch measure uses a file to determine the maximum depth of each WordNet taxonomy. The wup measure used a file to determine the depth of each synset in its taxonomy. The files used by these measures are produced by the wnDepths.pl program that is in the 'utils' directory.

The file used by lch has the version of WordNet that was used when generating the file on the first line of the file, similar to the information content files. For example:

 wnver::LL1BZMsWkr0YOuiewfbiL656+Q4

The remaining lines of the file give the part of speech for each taxonomy, the offset of the root of each taxonomy, and the maximum depth of the taxonomy. For example, the taxonomy that has the root entity#n#1 has this line:

 n 00001740 17

The file used by wup has also has the version of WordNet that was used when generating the file on the first line of the file. The remaining lines have the part of speech for the synset, the offset of the synset, and a list of depths and offsets of taxonomy roots. For example, the synset amphibious_landing#n#1, has the following line:

 n 00052540 4:00025950 4:00026194 5:00026194

The first item is the part of speech, the second is the offset of amphibious_landing#n#1. The next item, '4:00025950' means that there is a path from amphibious_landing#n#1 to the taxonomy root 00025950 (event#n#1) whose length is 4. The next item means that there is a path from amphibious_landing#n#1 to a different taxonomy root, 00026194 (act#n#2) whose length is 4. The last item means that there is another path to act#n#2 whose length is 5. From this we can infer that the synset belongs to two different taxonomies, and there there are multiple paths to the root of the second taxonomy.

SEE ALSO

intro.pod

Mailing list: http://groups.yahoo.com/group/wn-similarity

Project Home page: http://wn-similarity.sourceforge.net

AUTHORS

 Ted Pedersen, University of Minnesota, Duluth
 tpederse at d.umn.edu

 Siddharth Patwardhan, University of Utah, Salt Lake City
 sidd at cs.utah.edu

 Satanjeev Banerjee, Carnegie Mellon University, Pittsburgh
 banerjee+ at cs.cmu.edu

 Jason Michelizzi

COPYRIGHT AND LICENSE

Copyright (c) 2005-2008, Ted Pedersen, Siddharth Patwardhan, Satanjeev Banerjee, and Jason Michelizzi

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation; with no Invariant Sections, no Front-Cover Texts, and no Back-Cover Texts.

Note: a copy of the GNU Free Documentation License is available on the web at http://www.gnu.org/copyleft/fdl.html and is included in this distribution as FDL.txt.




Hosting generously
sponsored by Bytemark