/ 
UMLS-Similarity-1.21
River stage one • 1 direct dependent • 1 total dependent
/ UMLS::Similarity::jcn

NAME

UMLS::Similarity::jcn - Perl module for computing the semantic relatednessof concepts in the Unified Medical Language System (UMLS) using the method described by Jiang and Conrath (1997).

CITATION

@inproceedings{JiangC97,
 Author = {Jiang, J. and Conrath, D.},
 Booktitle = {Proceedings on International Conference 
              on Research in Computational Linguistics},
 Pages = {pp. 19-33},
 Title = {Semantic similarity based on corpus statistics 
          and lexical taxonomy},
 Year = {1997}
}

SYNOPSIS

use UMLS::Interface;
use UMLS::Similarity::jcn;

my $umls = UMLS::Interface->new(); 
die "Unable to create UMLS::Interface object.\n" if(!$umls);

my $jcn = UMLS::Similarity::jcn->new($umls);
die "Unable to create measure object.\n" if(!$jcn);

my $cui1 = "C0005767";
my $cui2 = "C0007634";

$ts1 = $umls->getTermList($cui1);
my $term1 = pop @{$ts1};

$ts2 = $umls->getTermList($cui2);
my $term2 = pop @{$ts2};

my $value = $jcn->getRelatedness($cui1, $cui2);

print "The similarity between $cui1 ($term1) and $cui2 ($term2) is $value\n";

DESCRIPTION

This module computes the semantic similarity of two concepts in the UMLS according to a method described by Jiang and Conrath (1997). This measure is based on a combination of using edge counts in the UMLS 'is-a' hierarchy and using the information content values of the concepts, as describedin the paper by Jiang and Conrath. Their measure, however, computes values that indicate the semantic distance between words (as opposed to their semantic similarity). In this implementation of the measure we invert the value so as to obtain a measure of semantic relatedness. Other issues that arise due to this inversion (such as handling of zero values in the denominator) have been taken care of as special cases.

The IC of a concept is defined as the negative log of the probabilty of the concept.

To use this measure, a propagation file containing the probability of a CUI for each of the CUIs from the source(s) specified in the configuration file. The format for this file is as follows:

C0000039<>0.00003951
C0000052<>0.00003951
C0000084<>0.00003951
C0000096<>0.00003951

A larger of example of this file can be found in the icpropagation file in the samples/ directory.

A propagation file can be created using the create-icfrequency.pl and the create-icpropagation.pl programs in the utils/ directory. The create-icfrequency.pl program takes plain text and returns a list of CUIs that are mapped to the text and the CUIs frequency counts. This file can then be used by the create-icpropagation.pl program to create a file containing a list of CUIs and their probability counts, or used directly by the umls-similarity.pl program which will calculate the probability of a concept on the fly.

SELF SIMILARITY

Since the Jiang and Conrath measure was initially calculated as a distance measure and turned into a similarity measure, we need to take care ofthe special cases in which the similarity of the two concepts results in zero but does not mean that the two concepts are not similar. Here is an explaination of how we did and why. This is taken from the discussion about this measure when it was being implemented in WordNet::Similarity. The actual message chain is located here:

http://tech.groups.yahoo.com/group/wn-similarity/message/8

The Jiang and Conrath measure is calculated as follows:

sim(c1, c2) = 1 / distance(c1, c2)

where

c1, c2 are the two concepts,
distance(c1, c2) = ic(c1) + ic(c2) - (2 * ic(lcs(c1, c2)))
ic               = the information content of the concept.
lcs(c1, c2)      = the least common subsumer of c1 and c2.

Now, we don't want distance to be 0 (=> similarity will become undefined). The distance can be 0 in 2 cases...

(1) ic(c1) = ic(c2) = ic(lcs(c1, c2)) = 0

ic(lcs(c1, c2)) can be 0 if the lcs turns out to be the root node (information content of the root node is zero). But since c1 and c2 can never be the root node, ic(c1) and ic(c2) would be 0 only if the 2 concepts have a 0 frequency count, in which case, for lack of data, we return a relatedness of 0 (similar to the lin case).

Note that the root node ACTUALLY has an information content of zero. Technically, none of the other concepts can have an information content value of zero. We assign concepts zero values, when in reality their information content is undefined (due to zero frequency counts). To see why look at the formula for information content: ic(c) = -log(freq(c)/freq(ROOT)) {log(0)? log(1)?}

(2) The second case that distance turns out to be zero is when...

ic(c1) + ic(c2) = 2 * ic(lcs(c1, c2))

(which could have a more likely special case ic(c1) = ic(c2) = ic(lcs(c1, c2)) if all three turn out to be the same concept.)

How should one handle this?

Intuitively this is the case of maximum relatedness (zero distance). For jcn this relatedness would be infinity... But we can't return infinity. And simply returning a 0 wouldn't work... since here we have found a pair of concepts with maximum relatedness, and returning a 0 would be like saying that they aren't related at all.

So what could we return as the maximum relatedness value?

So the way I handled this was to try to find the smallest distance greater than 0, so that sim would be a very high value, but not infinity. To find this value of distance I consider the formula of distance...

distance = ic(c1) + ic(c2) - (2 * ic(lcs(c1, c2)))

we get distance = 0 if ic(c1) = ic(c2) = ic(lcs(c1, c2)) So consider the case that ic(c2) = ic(lcs(c1, c2), but ic(c1) is the information content value just slightly more than that of ic(c2) (and ic(lcs(c1, c2))). We want to find the value of distance corresponding to such a case and this would be the next highest value of distance after 0.

We could select ic(c2) and ic(lcs(c1, c2)) to represent a highly specific concept or a highly general concept for this computation... We'll decide which one to select later... For now we want a formula to represent a value of distance = "almost zero".

ic(concept) = -log(freq(concept)/freq(root))

For ic(c1) to be just slightly more than ic(c2) (or ic(lcs(c1, c2))), what if we just reduced freq(concept) in the above formula by 1. i.e.

ic(c2) = ic(lcs(c1, c2)) = -log(freq/rootFreq)

ic(c1) = -log((freq-1)/rootFreq)

Since frequency is counted in whole numbers, this is the closest ic(c1) could be to ic(c2) (but not equal to it). With this formula we would have

distance = ic(c1) + ic(c2) - (2 * ic(lcs(c1, c2)))
         = ic(c1) + ic(c2) - (2 * ic(c2))

... since ic(c2) = ic(lcs(c1, c2))

= ic(c1) - ic(c2)
= -log((freq-1)/rootFreq) + log(freq/rootFreq)

Now comes the part where we want to decide whether to select a highly specific concept or a highly general concept for ic2 and ic3... I selected them to be the most general concepts for some non mathematical reasons (tho' I think I had come up with some mathematical ones)...

My reasons...

The most general concept is the root node... we always have the frequency count of the root node (non zero)... (if the root node is zero then there is something really wrong with the information content computed). It would be very difficult to find the most specific concept (tho' not impossible).

Somehow, mathematically, I had a feeling that the more general ic(c1) and ic(c2) are, they would be closer to each other on the log scale than if they were more specific concepts (I could be mistaken and it could be the other way around... and I don't have a proof right now to support what I'm saying)

anyway, taking the most general concepts (the root concept), we have

distance = -log((rootFreq - 1)/rootFreq) + log(rootFreq/rootFreq) = -log((rootFreq - 1)/rootFreq) + log(1) = -log((rootFreq - 1)/rootFreq)

This is the distance corresponding to "almost zero"... And this is what I put in the code for the 0 case (sim = infinity case).

With the hocus pocus above I have made an artificial bound on relatedness to "almost infinity".

PROPAGATION

The Information Content (IC) is defined as the negative log of the probability of a concept. The probability of a concept, c, is determine by summing the probability of the concept ocurring in some text plus the probability its decendants occuring in some text:

For more information on how this is calculated please see the README file.

USAGE

The semantic relatedness modules in this distribution are built as classes that expose the following methods: new() getRelatedness()

TYPICAL USAGE EXAMPLES

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

use UMLS::Similarity::jcn;
$measure = UMLS::Similarity::jcn->new($interface);

The reference of the initialized object is stored in the scalar variable '$measure'. '$interface' contains an interface object that should have been created earlier in the program (UMLS-Interface).

If the 'new' method is unable to create the object, '$measure' would be undefined.

To find the semantic relatedness of the concept 'blood' (C0005767) and the concept 'cell' (C0007634) using the measure, we would write the following piece of code:

$relatedness = $measure->getRelatedness('C0005767', 'C0007634');

CONFIGURATION OPTION

The UMLS-Interface package takes a configuration file to determine which sources and relations to use when obtaining the path information.

The format of the configuration file is as follows:

SAB :: <include|exclude> <source1, source2, ... sourceN>

REL :: <include|exclude> <relation1, relation2, ... relationN>

For example, if we wanted to use the MSH vocabulary with only the RB/RN relations, the configuration file would be:

SAB :: include MSH REL :: include RB, RN

or

SAB :: include MSH REL :: exclude PAR, CHD

If you go to the configuration file directory, there will be example configuration files for the different runs that you have performed.

For more information about the configuration options please see the README.

SEE ALSO

perl(1), UMLS::Interface

perl(1), UMLS::Similarity(3)

CONTACT US

If you have any trouble installing and using UMLS-Similarity, 
please contact us via the users mailing list :

    umls-similarity@yahoogroups.com

You can join this group by going to:

    http://tech.groups.yahoo.com/group/umls-similarity/

You may also contact us directly if you prefer :

    Bridget T. McInnes: bthomson at cs.umn.edu 

    Ted Pedersen : tpederse at d.umn.edu

AUTHORS

Bridget T McInnes <bthomson at cs.umn.edu>
Siddharth Patwardhan <sidd at cs.utah.edu>
Serguei Pakhomov <pakh0002 at umn.edu>
Ted Pedersen <tpederse at d.umn.edu>

COPYRIGHT AND LICENSE

Copyright 2004-2011 by Bridget T McInnes, Siddharth Patwardhan, Serguei Pakhomov, Ying Liu and Ted Pedersen

This library is free software; you can redistribute it and/or modify it under the same terms as Perl itself.