Intelligent Information Management, 2009, 1, 122-127
doi:10.4236/iim.2009.12018 Published Online November 2009 (http://www.scirp.org/journal/iim)
Word Sense Disambiguation in Information
Retrieval
Francis de la C. Fernández REYES, Exiquio C. Pérez LEYVA, Rogelio Lau FERNÁNDEZ
Instituto Superior Politécnico, José Antonio Echeverría, Marianao, Cuba
Abstract: The natural language processing has a set of phases that evolves from lexical text analysis to the
pragmatic one in which the author’s intentions are shown. The ambiguity problem appears in all of these
tasks. Previous works tries to do word sense disambiguation, the process of assign a sense to a word inside a
specific context, creating algorithms under a supervised or unsupervised approach, which means that those
algorithms use or not an external lexical resource. This paper presents an approximated approach that com-
bines not supervised algorithms by the use of a classifiers set, the result will be a learning algorithm based on
unsupervised methods for word sense disambiguation process. It begins with an introduction to word sense
disambiguation concepts and then analyzes some unsupervised algorithms in order to extract the best of them,
and combines them under a supervised approach making use of some classifiers.
Keywords: disambiguation algorithms, natural language processing, word sense disambiguation
1. Introduction
The natural language processing involves a set of tasks
and phases that evolves from the lexical text analysis to
the pragmatic one in which the author’s intentions are
shown. One natural language problem is ambiguity, as
we can see in the following sentence: “I made her duck”.
This is a classical example of ambiguity; someone who
hears this phrase understands the speaker’s intention, but
it is harder make the computer understands it. First, the
words duck and her are morphologically or syntactically
ambiguous in their part-of-speech. Duck can be a verb or
a noun, while her can be an object pronoun or a posses-
sive adjective. Second, the word make is semantically
ambiguous; it can mean create or cook. Finally, the verb
make is syntactically ambiguous in a different way.
Make can be transitive, that is, taking a single direct ob-
ject, or it can be intransitive, that is, taking two objects,
meaning that the first object (her) got made into the sec-
ond object (duck). Finally, make can take a direct object
and a verb, meaning that the object (her) got caused to
perform the verbal action (duck) [1].
There are already a lot of works that resolve, almost
complete, the lexical ambiguity, as part of syntactic
analysis in natural language processing. However, a cur-
rent problem, without a complete solution yet, is seman-
tically ambiguity, according to Alexander Gelbukh and
Grigori Sidorov [2]. Nowadays, algorithms that attempt
resolving this problem are divided into two groups: dis-
criminative algorithms and disambiguation algorithms.
One of the natural language processing applications is
the information retrieval. On the one hand, Internet and
digital libraries have a huge amount of knowledge that
can answer a lot of questions that people may have. On
the other hand, the amount of information is so huge that
interferes in its proficiency because it is impossible to
process it easily. At present, more used techniques for
information retrieval implicate the search of keywords:
Files that contain the words that the user indicates are
being found. Another idea to set relevant knowledge, in
front of a question, will be attending to synonym rela-
tions to establish similar documents; besides, the use of
relations between words under a specific context create
the necessity of employ a word sense disambiguation
algorithm to understand the sense of the words that user
is using in his search.
This paper presents an analysis of the existing disam-
biguation algorithms and it analyses the quality of each
of them taking into account the metrics that have been
establish for the evaluation. At the same time, it shows a
possible combination of features and classifiers to pro-
pose a word sense disambiguation algorithm that re-
solves some deficiencies detected before and improve the
evaluation parameters.
2. Word Sense Disambiguation Algorithms
The word-sense disambiguation process consists of as-
signing to each given word in a context, one definition or
meaning (predefine sense or not), that is distinguishable
from others that it can have.
The disambiguation techniques may be classified into

F. C. F. REYES ET AL. 123
the way that is shown in Figure 1.
In the case of discrimination algorithms, a meaning for
the word is not enough, it is necessary to determine
which occurrences have the same sense, without the need
of establish which it is. Besides, they work with no lin-
guistic resources. On the contrary, disambiguation algo-
rithms reach the meaning of the word using external lin-
guistic resources (corpus or knowledge bases). The dis-
crimination algorithms identify context vectors for all
given word occurrences, divide the vectors into groups
and interpret each of them as a sense.
Corpus based methods (supervised) collect a set of
examples, manually tagged, for each sense of disambi-
guation words and induce a classifier (Support Vector
Machine, Näive Bayes [3] from these examples, then the
disambiguation is reduced to the process of classifying
the word in one of his possible senses. Among the limita-
tions you can find in those methods we have the knowl-
edge domain dependency (due to the example set use)
and the manual tagging is extremely expensive.
Knowledge based methods (not supervised) do not
require tagged corpus, rather use external linguistic re-
sources, and therefore, the disambiguation can be con-
sidered at any knowledge domain, as long as that exter-
nal resource accept it. The supervised methods obtain
better results than the not supervised ones, but they pre-
fer the seconds because they are not restricted to a spe-
cific knowledge domain. Knowledge based methods can
make use of dictionaries as Lesk’s algorithm [4], of the-
saurus as Yarowsky’s [5] and of WordNet as Resnik’s
[6].
As not supervised algorithms can be applied to any
knowledge domain, as long as that external resource al-
lows it, the authors analyze just them in the rest of the
document. Specifically we will examine the following
methods: based on grouping, Fuzzy Borda Voting, Ex-
tended Lesk, Conceptual density and Sense probability.
There are others algorithms that also result interesting,
but we don’t analyze them in the present paper, examples
of these are: Meaning affinity model [7], using auto-
matically acquired predominant senses [8] and based on
lexical cohesion [9].
2.1. Using Sense Clustering for Word Sense
Disambiguation
This method doesn’t require the use of a training set, just
use WordNet. The algorithm begins to grouping all
senses of the disambiguation target words. This process
tries to identify cohesive senses’ groups; those are cre-
ated by means of Extended Star cluster algorithm and
0

-similarity graph between different senses [10]. Then,
the method filters the groups to select those that match
the best with the context. If the selected groups disam-
biguate all the words (each group show only one sense),
then the process stops and the senses belonging to the
selected groups are interpreted as the disambiguated ones.
Otherwise, the clustering and filtering steps are per-
formed again (regarding the remaining senses) until the
disambiguation is achieved or when it is impossible to
raise 0

threshold [10].
The algorithm input is the disambiguation target words
set W and the context represented as topic signature T
[10].
There are two sub process cluster and filter. The first
one is carried out by the Extended Star Clustering Algo-
rithm, which builds star-shaped and overlapped clusters.
Each cluster consists of a star and its satellites, where the
star is the sense with the highest connectivity of the
cluster, and the satellites are those senses connected with
the star. The connectivity is defined in terms of the 0

-
similarity graph, which is obtained using the cosine
similarity measure between topic signatures and the
minimum similarity threshold0

.
Once clustering is performed over the senses of words
in W, a set of sense clusters is obtained. As some clusters
can be more appropriate to describe the semantics of W
than others, they are ranked according to a textual mea-
sure context T.
WSD
From non predefined
senses (discrimination)
From predefined senses
(disambiguation)
Based on corpus
(supervised)
Based on Knowledge
(not supervised)
Figure 1. Classification of word sense disambiguation algorithms

F. C. F. REYES ET AL. 124
0
After grouping the algorithm is obtained a set S with
the senses of the selected group. Then, the method com-
pares if the sense quantity in S matches the word quan-
tity in W, if this is true, the disambiguation process stops,
on the contrary, the cluster algorithm starts again with
other threshold

.
The algorithm behaves adequately in front of nouns
instances, although it shows an inconvenience, it does
not disambiguate proper nouns due to the lack of senses
in WordNet. Although, it behaves in a non correctly way
for verbs, of 304 tested, the algorithm just disambiguated
the 30% in a correct way. This result is a consequence of
the high grade of polysemy that verbs had and the few
relations that appear between them in WordNet, therefore,
the cluster with high polysemy level can not create cohe-
sive groups and fail the disambiguation process. Besides,
the context is reduced to the sentence, conesquently, be-
come interesting extending the context to paragraph, this
approach includes more verbs and the search of WordNet
relations increases, the effect is the improvement of the
clustering process.
2.2. Combining Different Methods by Means
of Fuzzy Borda Voting
The original scheme, Borda voting score, was introduced
in 1770 by the mathematician Jean Charles de Borda of
the French Academy. This method consists of a ponder-
ing system, that fight the general believes that the candi-
date which obtains the majority of votes is the one that
voters prefer and show examples of contradictions in
votes’ system used before.
Borda voting score establishes a punctuation system
where each voter gives a mark to each one of the candi-
dates, following the order of their preferences. In order to
find the winner, add the obtained scores for each candi-
date. This system has an inconvenience it does not per-
form rules for the weights of the candidates, which could
imply an arbitrary assignment that changes the result.
The fuzzy variant allows the experts (voters on the
original scheme) gives a numerical value that indicates
how some alternatives (candidates on the original sche-
me) are preferred among others, evaluating the prefer-
ences in a range between 0 and 1. In Rosso & Buscaldi
[11], each expert gives a mark to each alternative, ac-
cording to the number of alternatives worse than it. The
algorithm establishes the use of experts (other disam-
biguation methods) to achieve the disambiguation proc-
ess, and the ones considered are: Sense probability, Ex-
tended Lesk, Conceptual density (for verbs) and Word-
Net domains.
To obtain fuzzy preferences relations, the output
weights of each expert k are transformed
to fuzzy confidence values by means of the following
transformation [11]:
12
,,..., n ww w
ki
ij
ij
w
rww 
k
ij
r
i
where is considered as the degree of confidence
with which the expert k prefers alternative
x
to j
x
.
With the fuzzy preferences relations of m experts over
n alternatives x1, x2, . . ., xn. For each expert k we obtain
a matrix of preference intensities [11]:
11 121
21 2221
kk k
n
kk k
rr r
rr r






1
() ()
m
ii
k
rx rkx


k
ij
r
12
kk
k
nnnn
rrr





The final value assigned by the expert k to each alter-
native xi coincides with the sum of the entries greater
than 0.5 in the i-th row in the preference matrix.
Therefore, the definitive fuzzy Borda count for an al-
ternative xi is obtained as the sum of the values assigned
by each expert k [11]:
In order to apply this system on word sense disam-
biguation the first thing to do is determine the target
word and the senses set of it, afterward, using an expert
(for example, “Sense probability”), calculate the weights
of each sense and begin the competition between them.
Establish the confidence degree of the first sense
with the second one, then with the third one, and so on in
order to make the competition. These fuzzy values are
part of the first row of the preferences matrix, each row
of the matrix match an analyzed sense. In this way are go
establishing the different confidence degrees with the
other senses till complete the matrix. To end with the
expert, establish the value assigned to the sense consid-
ered, this value is calculated adding the row that matches
with the alternatives proposed, but avoiding the value
founded in matrix principal diagonal because it is the
alternative analyzed with it self. Then proceed in same
way with the other experts, at the end add, for each sense,
the values that each one of them assigned and the biggest
score is the winner, considering this one as the correct
sense of the target word.
The algorithm has as principal idea the usage of a vot-
ing system with fuzzy bases where each expert is word
sense disambiguation algorithm knowledge based, these
is a positive issue, the approach put under competition
various algorithms and the correct sense is that one who
receive the biggest score. The ambiguity resolution for
verbs behaves adequately, the experts that are considered
work in the sentence context. Will be interesting find ano-

F. C. F. REYES ET AL. 125
ther set of experts, this selection should be doing ac-
cording to the algorithm results taking into account the
part of speech and considering also a baseline as expert,
then realize tests changing the competition scheme to
analyze the behavior of disambiguation process.
2.3. Sense Probability (Most Frequently Sense)
In nineties grow up an interest on the line of the estab-
lishment of methods for automatic word sense disam-
biguation evaluation that offers a reference point to
measure the quality of done work. A starter point is the
set of a superior and inferior limit for the disambiguation
results: the inferior limit match the choice of the most
frequently sense, while the superior limit match the hu-
man tagged. The authors asseverate as correct choice the
most frequently sense in a 75% of the cases. About supe-
rior limit, human tagged, exist many controversies about
his reliability, fed by contradictory results obtained in
several experiments. Other contribution is the morph-
syntactic and semantic tagged words of open class from
Brown corpus, making benchmarks for disambiguation
systems in senses choosing: the chance, the frequency
(most frequency sense), the concurrence [12].
Those basic disambiguation techniques, that usually
not imply any kind of linguistic knowledge, are used as a
reference point for the evaluation of those systems [12].
One of the baseline measure used is Sense probability.
It consists in assign the first sense of WordNet to each
target word and it is calculated as follow:
||
1
1(,)
||
T
MFS i
i
BLw k
T



(,)
i
wk where

is equal to 1 when the k-th sense of the
word belongs to the group manually tagged by the
lexicographer (example, SemCor) for word and it is 0
otherwise. The Most Frequency Sense (MFS) calcula-
tions are based on the frequencies of SemCor corpus
terms. T is the test corpus where are contained the in-
stances . WordNet rank the senses taking into account
the appearing frequencies of them inside SemCor corpus.
This baseline always resolves the semantic ambiguity
problem with a 78% precision [12].
i
w
i
w
i
w
This technique lack totally of linguistic information,
however, is a good idea has a tagged corpus to find the
most frequently sense. Actually, it is not a knowledge
based disambiguation algorithm but it is used as a refer-
ence point to compare the algorithm done.
2.4. Extended Lesk Algorithm
This algorithm is an improvement version (WordNet
based) of the well known Lesk procedure and based it on
the use of dictionaries. The original algorithm was sup-
ported in the comparison of the gloss target word with
the context words and their gloss. The improvement con-
sists in taking into account also the gloss of the concepts
related to the target word, by means of various WordNet
relations [13].
Then, the similarity between a word sense and his
context is calculated by overlapping. Overlapping is the
intersection between to synsets set (a synset is a number
that implies sense for WordNet), on one hand, the ones
for the target word and on the other hand the gloss of the
context synsets. To the target word is assigned the sense
obtained that betters overlap with the gloss of the context
words and their related synsets [13].
The heterogeneous approach of the Extended Lesk al-
gorithm show better results than the homogeneous ones,
consider the part of speech not improve the precision,
excluding the case of adjectives. The context of this al-
gorithm is a window of size 7, 9 or 11 around the target
word, does not take into account the sentence context
and it uses various WordNet relations and glosses. The
size of the window depends directly from the part of the
speech, while biggest is the window more decrease the
polysemy and reach better precision results, most of all
in verbs.
2.5. Conceptual Density Algorithm
Conceptual density was originally introduced by Agirre
and Rigau in 1996 above WordNet. It is calculated over
sub-graphs of this lexical database, determined by hy-
pernymy relations. The proposed formula gives a meas-
ure of the conceptual density between the word and their
senses. Besides, this formula has the following charac-
teristics [14]:
 The length of the shortest path that connects the
concepts involved.
 The independent measure of the number of con-
cepts we are measuring.
 The depth in the hierarchy: concepts in a deeper
part of the hierarchy should be ranker closer.
Given a concept c, at the top of a sub-hierarchy, and
given nhyp and h (mean number of hyponyms per node
and height of the sub-hierarchy, respectively), the Con-
ceptual Density for c when its sub-hierarchy contains a
number m (marks) of senses of the words to disambigu-
ate is given by the formula below [14]:
0.20
1
0
(, )
m
i
i
c
nhyp
CD c mdescendientes



Rosso and Buscaldi [11] transform this conception of
Agirre and Rigau to use the formula as an expert in the
Fuzzy counting Borda system. The formulation of Con-
ceptual Density for a sub-graph S of WordNet is defined
by the following formula:

F. C. F. REYES ET AL. 126
Table 1 shows a comparison between the not super-
vised word sense disambiguation algorithms according to
the previous metrics [12]:
(,,)( )
m
CD mfnmn


where m is the relevant synsets in the sub-graph and n is
the total number of them. The constant
Even when Sense probability, using always the most
frequently sense, is the one that occupies the first place,
has a limitation, do not take into account linguistic infor-
mation. Therefore, the best systems are those that use
clustering and Fuzzy Borda counting. It is part of sup-
posing then that the combination of them offer better
results.

takes values
around 0.1, and this value was obtained through experi-
mental results from Rosso and colleagues work [15]. The
argument value f is obtained from the frequency of the
synsets related in the sub-graph, which appears in
WordNet. The relevant synsets match with the ones of
the target word and the ones of the context words. At
Buscaldi and Rosso [11] the expert that perform over
Conceptual Density use as context two nouns for the
word sense disambiguation process. The weights that are
calculated by the previous formula are used to compute
the confidence values used for fill the preference matrix.
It proposes, besides, a second expert that exploits the
holonymy relations instead of hypernymy. This expert
uses as context all nouns in the sentence where appears
the target word [11].
4. Word Sense Disambiguation Algorithm
Propose
Supervised methods offers better solution than not su-
pervised ones, however, the first ones need a training
tagged set, this implies human factor for tagging and
lacks a representative corpus for Spanish. Not supervised
methods, on the other hand, do not require a training
corpus, they use eternal resources as lexical data bases,
thesaurus, and dictionaries, those are available in Internet
and they abound most of all for the English language.
Buscaldi and Rosso proposition reach good precision
(around 75%) over nouns and use a window of two
nouns. For verbs does not behaves in a adequately way,
taking into account Banerjee analysis [13] we can cor-
roborate that to resolve correctly the verb ambiguity it is
necessary use windows of at least 11 context words and
besides including the major quantity of linguistic infor-
mation.
There are propositions that try to combine not super-
vised methods by a voting system [11,15] obtaining good
results. It would be interesting combine not supervised
methods under a supervised approach, that means, make
a proposition where the first step is the selection of algo-
rithm set taking into account their results in the word
sense disambiguation process and then apply to them a
classifier set to learn which method has to use under a
determined circumstance. Previously we analyze some
algorithms propositions that show satisfactory results and
we conclude that ambiguity on verbs is harder to result,
that is why it makes necessary to include as much lin-
guistic information as it is possible and enlarge the sen-
tence context to paragraph context.
3. Comparison of the Word Sense
Disambiguation Algorithms
As it is known, the metrics that are used to evaluate a
disambiguation algorithm are the following [12]:
Precision= # correctly disambiguated words
# disambiguated words
Recall = # correctly disambiguated words The method we propose is based on the combination
of various not supervised algorithms and baselines. It
creates a feature vector (some features are obtained from
WordNet) for each sense of the target word. Each algo-
rithm is a feature for each sense, the output of the algo-
rithm will be normalized in a measure between 0 and 1,
then it combines certain classifiers to search the winner
sense. If no exit one, then it uses a baseline (for instance
MFS) in order to search the correct sense of the word.
# tested set words
Coverage = # disambiguated words
# tested set words
The combination of precision and recall it is known as
F1 measure and it is calculated by the following formula
[12]:
F1 = 2 * precision * recall
(precision + recall)
Table 1. Not supervised system score ranked by F1 measure. (C=Coverage, P=Precision, R=Recall, F1=F1 measure)
System C P R F1
M
FS
BL 100.0 78.89 78.89 78.89
Fuzzy Borda 100.0 78.63 78.63 78.63
Clustering based 100.0 70.21 70.21 70.21
Conceptual density 86.2 71.2 61.4 65.94
Extended Lesk 100.0 62.4 62.4 62.4

F. C. F. REYES ET AL. 127
The method defines a target word as a set of sense
vectors. Each vector contains in the beginning all kind of
linguistic information and the normalized result of ap-
plying some algorithms (the first experiments will use
Extended Lesk, Conceptual Density and Clustered based
algorithms). Then the vectors are reduced by use of fea-
ture selection eliminating linguistic redundancies among
vectors. To find the winner sense we suggest the use of
the combination of algorithms results that we found on
each feature vector, then apply a classification technique
such as Support Vector Machines or Näives Bayes which
performs good over two class, this classes proposed are
best and bad senses. We hope this method increase the
precision in word sense disambiguation process.
5. Conclusions
The supervised methods offer better solutions than not
supervised ones, but they do not make use of external
resources, that is why they are applied on specific do-
mains, using language characteristics and syntactic reso-
lution, making them language dependents.
The algorithm proposed tries to combine positive fea-
tures of the not supervised method, establishing a classi-
fier system to determine which of them is better, and
taking into account the winner algorithm, the proposition
output the correct sense of the word. Even when this
method is in a development phase and determination of
the classifiers to use, in order to validate it in some task
of the SemEval 2007 competition, empirically is well
suppose that this proposition shows better results taking
into account the enlargement of the considered linguistic
information in the algorithm previously analyzed and
they are combined under a automatic learning approach.
REFERENCES
[1] E. Agirre and G. Rigau, “Word sense disambiguation
using conceptual density,” International Conference on
Computational Linguistics (COLING), Copenhagen,
Denmark 1996.
[2] H. Anaya-Sánchez, A. Pons-Porrata, et al., “TKB-UO:
Using sense clustering for WSD,” 4th International
Workshop on Semantic Evaluations (SemEval), Prague,
Czech Republic, Association for Computational Linguis-
tics, 2007.
[3] S. Banerjee, “Adapting the lesk algorithm for word sense
disambiguation using wordnet,” Department of Comput-
ing Science. Minnesota, USA, University of Minnesota,
MSc.: 98, 2002.
[4] D. Buscaldi and P. Rosso, “UPV-WSD: Combining dif-
ferent WSD methods by means of fuzzy borda voting,”
4th International Workshop on Semantic Evaluations
(SemEval), Prague, Czech Republic, Association for
Computational Linguistics, 2007.
[5] Y. Chali and S. R. Joty, “UofL: Word dense disambigua-
tion using lexical cohesion,” 4th International Workshop
on Semantic Evaluations (SemEval), Prague, Czech Re-
public, Association for Computational Linguistics, 2007.
[6] A. Gelbukh and G. Sidorov, “Procesamiento automático
del español con enfoque en recursos léxicos grandes,”
México, Centro de Investigación en computación,
Instituto Politécnico Nacional, 2006.
[7] J. Huang, J. Lu, et al., “Comparing naive Bayes, decision
trees, and SVM with AUC and accuracy,” Third IEEE In-
ternational Conference on Data Mining, 2003.
[8] R. Ion and D. Tufis, “RACAI: Meaning affinity models,”
4th International Workshop on Semantic Evaluations
(SemEval), Prague, Czech Republic, Association for
Computational Linguistics, 2007.
[9] D. Jurafsky and J. H. Martin, “Speech and language
processing: An introduction to natural language process-
ing, computational linguistics, and speech recognition,”
Prentice Hall, 2000.
[10] R. Koeling and D. McCarthy, “Sussx: WSD using auto-
matically acquired predominant senses,” 4th International
Workshop on Semantic Evaluations (SemEval), Prague,
Czech Republic, Association for Computational Linguis-
tics, 2007.
[11] M. Lesk, “Automatic sense disambiguation using ma-
chine readable dictionaries: How to tell a pine cone from
an ice cream cone,” 5th Annual International Conference
on Systems Documentation, Ontario, Canada, ACM,
1986.
[12] I. Nica, “El conocimiento lingüístico en la desambi-
guación semántica automática,” España, 2006.
[13] P. Resnik, “Disambiguating noun groupings with respect
to wordnet senses,” Third Workshop on Very Large Cor-
pora, Massachusetts Institute of Technology Cambridge,
Massachusetts, USA, 1995.
[14] D. Yarowsky, “Word-sense disambiguation using statisti-
cal models of roget’s categories trained on large corpora,”
15th International Conference on Computational Linguis-
tics (COLING), Nantes, France, Association for Compu-
tational Linguistics, 1992.
[15] D. Yarowsky, “Unsupervised word sense disambiguation
rivaling supervised methods,” 33th Annual Meeting of
the Association for Computational Linguistics (ACL’95),
Cambridge, Massachussets, 1995.