Journal of Software Engineering and Applications, 20 12, 5, 175-180
doi:10.4236 /js ea.2012.512b034 Published Online December 2012 (
Copyright © 2012 SciRes. JSEA
Using Wikipedia as an External Knowledge Source for
Supporting Contextual Disambiguation
Shahida Jabeen, Xiaoying Gao, Peter Andreae
School of Engineering and Computer Science, Victoria University of Wellington, New Zealand.
Email: shahidar, Xiaoying. Gao@ ecs .vuw.a,
Received 2012
Every term has a meaning but there are terms which ha ve mult iple mea ning s. Ident ifyin g the cor rect mea ning o f a term
in a specific context is the goal of Word Sense Disambiguation (WSD) applications. Identifying the correct sense of a
ter m give n a l i mited c ont e xt i s eve n har d er . Thi s research aims at solving the problem of identifying the correct sense of
a ter m give n only one term as its conte xt. The main focus of this research is on using Wikipedia as the external know-
ledge source to decipher the true meaning of each term using a single term as the context. We experimented with the
semantically rich Wikipedia senses and hyperlinks for context disambiguation. We also analyzed the effect of sense
filtering on context extraction and found it quite effective for contextual disambiguatio n. Results have sho wn that dis-
ambiguatio n with filtering works quite wel l on manually disambiguated dataset with the performance accuracy of 86%.
Keywords: Contextual Disambiguatio n; Wikipedia Hyperlinks; Se mantic Related ne ss
1. Introduction
Ambiguity is implicit to natural languages with a large
number of terms having multiple contexts. For instance,
the English noun “bank" can be a financial institution or
it could mean river side. Polysemy is the capability of a
term to have multip le meanings and a term with multip le
senses is termed as “polysemous term”. Resolving am-
biguity and identifying the most appropriate sense is a
critical issue in natural language interpretation and
processing. Identification of the correct sense for an am-
biguous term requires understanding of the context in
which it occurs in natural language text. Word Sense
Disa mbiguation is defined as the task of automatically
selecting the most appropriate sense of a term within a
given context fro m a set of available senses [1-4].
What is the co mmo n cont ext of B ank a nd floo d? What
sense of Tree comes to mind when it occurs with com-
puter? In both cases, the relationship of each pair of
terms is strong but can’t be measured unless considering
the correct senses of either or both terms. The first pair of
terms has a cause and effect relation provided the correct
context of bank i.e Levee is taken into account. Whereas,
to relate the other term pair, the term Tree should be
mapped to Tree (data structure). Making judgments
about deciphering the relevant context of a term is an
apparently ordinary task but actually requires a vast
amount of background knowledge about the concepts.
The task of contextual information extraction for dis-
ambiguatio n of natur al language te xt relies o n knowledge
from a broad range of real world concepts and implicit
relations [5]. To effectively perform such a task, com-
puters ma y r eq ui re an external knowledge source to infer
implicit relations. External knowledge sources can vary
from domain specific thesauri to lexical dictionaries and
from hand crafted taxonomies to knowledge bases. Any
external knowledge source should have the following
three characteristics to sufficiently support contextual
disambi guation:
Coverage: The coverage of the external knowledge
source is the degree to which it represents collected
knowledge. It should be broad enough to provide in-
formation about all relevant concepts.
Quality: Information provided by the external know-
ledge source should be accurate, authentic and up-
Lexical semantic resource: The knowledge base
should encode rich lexical and structural semantics.
Fortunately, Wikipedia, a web based freely available
encyclopedia, comes up with all the necessary characte-
ristics to support contextual information extraction and
disambi guation. As a knowledge source, Wikipedia not
only focuses on general vocabulary but also covers a
large number of named entities and domain specific
terms. The specialty of Wikipedia is that each of its ar-
ticles is dedicated to a single topic with an additional
bene fit of heavy li nkin g between articles. Wikipedia also
Using Wikipedia as an External Knowledge Source for Supporting Contextual Disambiguation
Copyright © 2012 SciRes. JSEA
covers specific senses of a term, surface forms, spelling
variations, abbreviations and derivations. Importantly, it
has a semantically rich hyperlink network relating ar-
ticles that cover different types of lexical semantic rela-
tions such as hyponymy and hypernymy, synonymy, an-
tonymy and polysemy. Hence, Wikipedia sufficiently
demonstrates all the necessary aspects of a good know-
ledge source. It provides a knowledge base for extracting
information in a more structured way than a search en-
gine and with a coverage better than other knowledge
sources [6].
The main goal of this research is to use Wikipedia as
the external knowledge source to decipher true meaning
of a term using the other term as the context. In particular,
we have taken into account Wikipedia hyperlinks struc-
ture and senses for disambiguating the context of a term
pair. The rest of the paper is organized as follows. Sec-
tion 2 discusses two broad categories of disambiguation
work proposed in literature and the corresponding re-
search done in each category. Section 3 discusses our
proposed approach for context extraction and disambigu-
ation. Sectio n 4 comprises of the perfor mance analysis o f
our proposed methodology using a manually designed
dataset consisting of English term pairs. Finally, section
5 concludes our research and discusses some future re-
search directions.
2. Related Work
Word Sense Disambiguation (WSD) is a well explored
area with lot of research work still going on in context
identification and disambiguation. Research in this area
can be broadly categorized into two main streams [7]:
Lexical approaches and taxonomic approaches.
Lexical approaches are based either on the analysis of
a disambiguated corpus or by extracting strings from the
definition of that sense. Such approaches are based on
identifying the lexical features such as occurrence statis-
tics or co-occurrence computations. By contrast, tax-
onomic approaches identify correct nodes in the hie-
rarc hy of sense s and explore the relations between nodes
for c omputing t he semantic closeness.
Early efforts in Lexical approaches were based on
machine readable dictionaries and thesauri, associating
word senses with short definitions, examples or lexical
relations. A simple approach of this type is based on
comparing dictionary definitions of words, also called
glosse s, to the co ntex t words (the words appearing in the
surrounding text) of an ambiguous word. Clearly, the
higher the overlap between context words and the dic-
tionary definitio ns of a particular sense of the ambiguous
word, the higher the chances of getting the correct sense
for the word. Cowie et al. [8] and Lesk [9] based their
approaches on machine readable dictionaries and thesauri.
Pederson et al. [10] adapted the Lesk algorithm for word
sense disambiguation by using lexical database WordNet
instead of standard dictionary as a source of glosses.
They exploited the hierarchy of WordNet semantic rela-
tions for disambiguation task. Patwardhan et al. [1] ge-
neralized adapted Lesk algorithm to a method of term
sense disambiguation. They used the gloss overlap as an
effective measure of semantic relatedness. Pedersen [11]
further explored the use of similarity measures based on
path findings in concept networks, information contents
derived from large corpus and term sense glosses. They
concluded that the gloss based measures were quite ef-
fective for term sense disambiguation. Yarowsky [12]
used Statistical models on Roget’s thesaurus categories
to build context discriminators for the word senses that
are members of conceptual classes. A conceptual class
such as “machine” or “animal” tends to appear in recog-
nizably different contexts. They also used the context
indicators of Roget’s thesaurus as the context indicators
for the members of conceptual categories.
Following taxonomic approaches, Agirre [13] pro-
posed a method of lexical disambiguation over Brown’s
Corpus using no un taxo nomy of Word Net. He computed
the conceptual density by finding the combination of
senses from a set of contiguous nouns that maximized
conceptual distances in taxonomic concepts. Veronis et
al. [14] automatically build a very large Neural Network
from definitio n text in machine readable dictionaries and
used this network for word sense disambiguation. They
used the node activation scheme for moving closer to the
most related senses following the “Winn er-take-all”
strategy in which every active node sent an activation to
another increasingly related node in the network. Mihal-
cea et al. [15] used the page ranking algorithm on Se-
mantic Networks for sense disambiguation. Iterative
graph-based ranking algorithms are essentially a way of
deciding the importance of a node within a graph. When
one node in a graph is connected to another one, it is
casting a vote for that other node. The higher the number
of votes of a node , the higher the importance of that
node. They find the sysnet with highest PageRank score
for each ambiguous word in the text as suming that it will
uniquely identify t he sense of the word.
Wikipedia and Sense Disambiguation
Availability of free online thesaurus, dictionaries and
encyclopedias and other knowledge sources has scaf-
folded the improvement in both lexical and taxonomic
based word sense disambiguation. Different methodolo-
gies are found in literature for computing term sense
disambiguation based on Wikipedia as the external
knowledge source. Ponzetto and Navigli [4] addressed
the problem of knowledge acquisition in term sense
disambiguation. They proposed a methodology for ex-
Using Wikipedia as an External Knowledge Source for Supporting Contextual Disambiguation
Copyright © 2012 SciRes. JSEA
tending Wor dNet with large amount of semantic relatio ns
derived from Wikipedia. They associated Wikipedia
pages with the WordNet senses to produce a richer lexi-
cal resource. Rada [2] used Wikipedia as a source of
sense annotation for generating sense-tagged data for
building accurate and robust sense classifiers. Turdakov
and Velikhov [16] proposed a semantic relatedness
measure based on Wikipedia links and used it to disam-
biguate terms. They proposed four link-based heuristics
for r educ ing t he searc h space of potentially related topics.
Fogarolli [17] used Wikipedia as a reference to obtain
lexicographic relations and combined them with the sta-
tistical information to deduce concepts related to terms
extracted from a corpus. Cucerzan [18] presented an ap-
proach for the recognition and semantic disambiguation
of named entities based on agreement between informa-
tion extracted from Wikipedia and the context of Web
search results. Bunescu [19] addressed the same problem
of detecting and disambiguating named entities in open
domain text using Wikipedia as the external knowledge
The kind of pro blem that we add ress in this research is
a variant of the main stream of term sense disambigua-
tion research, where the aim is to identify the context of a
single term. We look to find o ut the co ntext of two ter ms
with respect to each other. This task is critical in many
approaches involving r elatedness computation [6,20, 21].
3. Disambiguation Methods
There are two approaches that we have adopted for con-
textual extraction and disambiguation along with their
variants based on two factors: relatedness measure and
sense filte rin g.
3.1. WikiSim Based Disambiguation
Our main approach for contextual disambiguation using
Wikipedia consists of three phases: Context Extraction,
in which we extract the candidate context of both input
terms ; Context Filtering, where we filter o ut certain co n-
texts based on t heir t ype, th us avo iding u nnece ssar y con-
text; Contextual Disambiguation, where semantic simi-
larity between candidate contexts is computed using Wi-
kipedia hyperlink based
1) Context Extraction: We start with identification of
all possible contexts corresponding to each input term
based on Wikipedia senses. Each Wikipedia article is
associated with a number of senses. For instance, there
are various senses for terms Present and Tense. T he best
sense of Present would be Present tense and the best
sense of tense would be Grammatical tense in correct
context of each other. The aim of this phase is to extract
all possible co ntexts corr espond ing to the inpu t ter m pair.
For this purpose, we used Wikipedia disambiguation
pages to extract all listed senses as candidate senses and
populate them in the conte xt set of each inp ut term.
2) Context Filtering: There are three broad categories
of manually annotate d Wikipedia senses.
Senses with parenthesi s
Single term senses
Phrasal senses
The first type of senses are those Wikipedia senses
which are generic and include broader context within
parenthesis followin g the title. Fo r exa mple, Crane (B ird )
and Crane (machine) are two different contexts of the
term Crane. Both of these senses are quite distinctive and
cover broader contexts of machine and bird .
In the second category of senses, single term context
falls. These contexts cover certain very important lexical
relations such as synonymy, hypernymy and hyponymy
and derivations. For instance, Gift is a synonym of the
term present, similarly carnivores is the hypernym of the
ter m tiger . I t is fou nd tha t s uc h sense s a re ver y use ful for
contextual information extr act ion.
The third type of Wikipedia senses are phrasal senses,
which usually have very specific and limited context.
This context can have characteristics of the more general
sense but it would be focused more on some other spe-
cific features which cannot be considered true for the
generalized sense. For instance, corresponding to a term
forest, the phrasal sense might be Forest Township, Mis-
saukee County, Michigan which discusses a geographi-
cally specific context rather than the more general con-
text of forest. Such type of senses might not be ver y u se-
ful in extracting the contextual information. For this rea-
son we excluded this third type of senses from the con-
text set of each input term. Experiments proved that bi-
gram senses still contain the general context of a term. So,
we gathered all uni-gra m and b i-gra m senses, se nses wit h
parenthesis and senses shared by both input terms and
put them in the context set corresponding to each input
3) Contextual Disambiguation: In contextual disam-
biguation, the first step is to extract all uniq ue inlinks (all
articles referring to the input term article) and outlinks
(all articles referred by the input term article) corres-
ponding to each candidate sense of the context set. The
link vector of each candidate sense of first input term is
compared with that of each sense of the second input
term. The assumption behind this comparison is to find
out those se nses which share maximum number of links,
thus indicating a strong relatedness. Each sense pair is
assigned a weight based on a relatedness measure called
WikiS im [20] , as shown below:
In the above formula si| Swa |, where |Swa | i s the set
Using Wikipedia as an External Knowledge Source for Supporting Contextual Disambiguation
Copyright © 2012 SciRes. JSEA
of all senses of term wa whereas, sj sense corresponds to
input term wb and si| Swb|. S is the set of all the links
shared between a sense pair and T represents total num-
ber of links of both senses. In other terms, the weight of a
sense p air is the link probability of shared links, or is 0 if
shared links do not exist. Onc e we get scores for all sense
pair s, the next step is to find out the sense pair with
highest score, thus getting most closely related senses of
both input terms. For this purpose, all sense pairs are
sorted based on their WikiSim score and the sense pair
having the highest weight is taken as the disambiguated
context corresponding to the input term p a ir .
3.2. WLM based Disambiguation
In or d er to e va l ua t e t he pe r fo r ma nc e o f our ma i n a pproach
and to analyze the effect of using a different relatedness
measure on disambiguation task, we used WLM related-
ness measure [21] which is also based on Wikipedia
hyperlinks and is proven to be quite effective in compu-
ting term relatedness. We followed the same methodolo-
gy as the WikiSim based Disambiguation for extracting
candidate senses and populating the context vector but
replaced the WikiSim relatedness measure with WLM
measure while computing the sense pair scores.
4. Evaluation
4.1. Experimental setup
We used the version of Wikipedia released in July,
20111. At this point, it contains 31GB of uncompressed
XML markup which corresponds to more than five mil-
lion articles which sufficiently covers all the co ncepts for
Table 1. Accuracy Based Performance Comparison Of
Disambiguation M e thods.
Table 2. Statistics Of Types Of Disambiguation Performed
By Each Metho d.
which manual judgment were available. To explore and
easily draw upon the contents of Wikipedia, we used the
latest version (wi kiped ia-miner-1.2.0) of Wikiminer
toolkit [22] which is an open source Java code2. Since
the problem addressed in this research is a variant of the
standard disambiguation task, where rather than resolv-
ing the context of a single term we do that for a pair of
terms considering each word as the context for the other
word, we needed a different dataset of disambiguated
term pairs. So, we used a manually designed dataset
named WikiContext as shown in Table 3. It consists of
30 E nglish ter m pair s which ar e manual ly disa mbi guated
to corresponding Wikipedia articles in context of the
other input term.
4.2. Experimental Results and Discussion
To compare performance of our proposed methods, we
automatically disambiguated term pairs in the dataset and
compared them with the manually disambiguated Wiki-
pedia articles. To measure the performance of each me-
thod, we used the accuracy measurement:
thod defined as follow:
where, |Pc| is the set of correctly disambiguated ter m
pairs and |Pt| refers to the set of all disambiguated term
pairs. In other words, it is the ratio o f correct disambigu-
ation and t he size of the dataset.
Table 3. Wikisim (With Filtering):Word Pairs And Cor-
respo nding Aut omatic Disambiguation.
Using Wikipedia as an External Knowledge Source for Supporting Contextual Disambiguation
Copyright © 2012 SciRes. JSEA
When compared performance accuracy of both Wiki-
Sim based disambiguation and WLM based disambigua-
tion, WikiSim based disambiguation is found to be com-
parable to that of WLM based disa mbiguation, both hav-
ing an accuracy of 86% as shown in Table I. To analyze
the effect o f applying sense filterin g in both disambigua-
tion approaches, we skipped context filtering step from
each approach and performed disambiguation with all
possible contexts. Detailed analysis of each approach is
summarized in Table II. Three types of disambiguation
are taken into account in this research: First, those word
pairs which are matched exactly to the manual disam-
biguat ion; se co nd , those word pairs which are matched to
a specialized area or subtopic of the correct context; third,
those word pairs in which one of the term is correctly
disambiguated in context of the other word but the other
term is disambiguated to a wrong context. Results of
WikiSim based disambiguation revealed that there was
no partial match in case of both context filtering and
without filtering. Whereas, some of the specialized
matches were found to be more relevant then the exact
matches. For example in case of the term pair <mole,
chemistry>, mole was disambiguated exactly to Mole
(unit) which is the measurement of amount of chemical
substance but che mistry was disambi guated to even mor e
related context Analytical Chemistry which deals with
quantification of the chemical components. Similarly, the
term pair <bar, drinking> was disambiguated to <Bar
(Establishment), Alcoholic beverages>. In case of WLM
based disambiguation with filtering, three term pairs
were found to be partially matched to the correct context.
Table III shows the results on WikiSim based disambig-
uation (with filtering) on the dataset WikiContext. It
shows disambiguated Wikipedia articles corresponding
to input term pairs.
Table I
Accuracy Based Performance Comparison Of Disam-
biguation Method s
Overall, disambiguation based on filtering is found to
be better than the one without filtering. The accuracy of
WikiSim based method is 3% increased when filtering is
applied. The effect of filtering became more evident in
WLM where the accuracy of filtering based disambigua-
tion increased to 10%. The results of our main approach
are quite encouraging and comparable to the WLM based
disambiguation. In order to avoid any bias in the results
due to smaller size of dataset and to test the effectiveness
of our approach more critically, we plan to use some
bigger dataset in future. To the best of our knowledge,
there is no dataset available that addresses this particular
kind of pro blem. One li mitati on of our approach is that it
relies heavily on Wikipedia senses, which are manually
encoded and may not sufficiently cover all possible con-
texts of so me terms and suffers from inco nsiste nt for mat-
ting d ue to manua l e ncod ing. We bel ieve tha t usi ng o ther
Wikipedia features such as anchor text s , categories,
hyperlinks and redirects for semantic extraction would
definitely help in t his regard.
5. Conclusion
In this paper we proposed and evaluated a novel approach
for extracting contextual information from Wikipedia
and using it to d isambi guate a term using a single ter m as
a given context. Based on Wikipedia hyperlink structure
and senses, our approach used WikiSim, a Wikipedia
based relatedness measure to compute scores of sense
pairs and compare them based on their relatedness. For
the sense disambiguation, we extracted various senses of
first input term and disambiguated each sense in various
contexts of the other input term. We evaluated the per-
formance of our approach by comparing it with WLM
based disambiguation approach and analyzed the effect
of context filtering disambiguation. Results have shown
that with an accuracy of 86%, our approach performs
quite well when compared with manually disambiguated
dataset of term pairs. In future, we plan to apply this
disambiguation approach along with the semantic rela-
tedness in key phrase clustering task for an indirect
evaluation of our approach on a bigger dataset to avoid
any bias in the current results due to smaller dataset size.
[1] S. Patwardhan, S. Banerjee, and T. Pedersen, “Using
measures of semantic relatedness for word sense disam-
biguation,” in Proceedings of the 4th International Con-
ference on IntelligentText Processing and Computational
Linguistics, February 2003, pp. 241257.
[2] R. Mihalcea, “Using wikipedia for automatic word sense
disambiguation,” in North American Chapter of the As-
sociation for Computational Linguistics (NAACL 2007),
[3] D. McCarthy, “Word sense disambiguation: The case for
combinations of knowledge sources,” Natural Language
Engineering, vol. 10, pp. 196–200, June 2004.
[4] S. P. Ponzetto and R. Navigli, “Knowledge-rich word
sense disambiguation rivaling supervised systems,” in
Proceedings of the 48th Annual Meeting of the Associa-
tion for Computational Linguistics, 2010, pp. 1522–1531.
[5] E. Yeh, D. Ramage, C. D. Manning, E. Agirre, and A.
Soroa, “Wikiwalk: random walks on wikipedia for se-
mantic relatedness,” in 2009 Workshop on Graph-based
Methods for Natural Language Processing, 2009, pp.
[6] M. Strube and S. P. Ponzetto, “Wikirelate! computing
semantic relatedness using wikipedia,” in proceedings of
the 21st national conference on Artificial intelligence, vol.
Using Wikipedia as an External Knowledge Source for Supporting Contextual Disambiguation
Copyright © 2012 SciRes. JSEA
2, 2006, pp. 1419–1424.
[7] J. C urtis, J. Cabral, and D . Baxter, “On t he appli cation of
the cyc ontology to word sense disambiguation,” in Pro-
ceedin gs of the 19th International Florida Artificial Intel-
ligence Research Society Conference, 2 006, pp. 652657.
[8] J. Cowie, J. Guthrie, and L. Guthrie, “Lexical disa mbigu-
ation using simulated annealing,” in Proceedings of the
workshop on Speech and Natural Language, 1992, pp.
[9] M. Lesk, “Automatic sense disambiguation using ma-
chine readable dictionaries: how to tell a pine cone from
an ice cream cone,” in Proceedings of the 5th annual in-
ternati onal conference on Systems documentation, 1986,
pp. 24–26.
[10] S. Banerjee and T. Pedersen, “An adapted lesk algorithm
for word sense disambiguation using wordnet,” In Pro-
ceeing of the Third International Conference on Intelli-
gent Text Processing and Computational Linguistics,
2002, pp . 136–145.
[11] T. Pedersen, S. Banerjee, and S. Patwardhan, “Maximiz-
ing Semantic Relatedness to Perform Word Sense Dis-
ambiguation,” University of Minnesota Supercomputing
Institute, Research Report UMSI 200 5/25, Mar c h 20 05.
[12] D. Yarowsky, “Word-sense disambiguation using statis-
tical models of roget’s categories trained on large corpo-
ra,” in Proceedings of the 14th conference on Computa-
tional linguistics - Volume 2, 1992, pp. 454460.
[13] E. Agirre and G. Rigau, “Word sense disambiguation
using conceptual density,” in Proceedings of the 16th
conference on Computational linguistics - Volume 1,
1996, pp . 1622.
[14] J. Veronis and N. M. Ide, “Word sense disambiguation
with very large neural networks extracted from machine
readabl e dictionaries,” in Proceedings of the 13th confe-
rence on Computational linguistics - Volume 2, 1990, pp.
[15] R . Mihalcea, P. Tarau, and E. Figa, “P agerank on se man-
tic networks, with application to word sense disambigua-
tion,” in Proceedings of the 20th international conference
on Computational Linguistics, 2004.
[16] D. Turdakov and P. Velikhov, “Semantic relatedness
metric for wikipedia concepts based on link analysis and
its application to word sense disambiguation,” SYRCo-
DIS, vol . 35 5, pp. 16, 2008.
[17] A. Fogarolli, “Word sense disambiguation based on wi-
kipedia link structure,” in Proceedings of the 2009 IEEE
International Conference on Semantic Computing, 2009,
pp. 77–82.
[18] S. Cucerzan, “Large-scale named entity disambi g uation
based on wikipedia data,” in Proceedings of EMNLP-
CoNLL 2007, 2007, pp. 708–716.
[19] B. Razvan and P. Marius, “Using encyclopedic know-
ledge for named entity disambiguation,” in Proceesings of
the 11th Conference of the European Chapter of the Asso-
ciation for Computational Linguistics (EACL-06), 2006,
pp. 916.
[20] S. Jabeen, X. Gao, and P. Andreae, “Improving contex-
tual relatedness computation by leveraging wikipedia se-
mantics, ” in 12th Paci fic R im International Conference on
Artificial Intelligence (To appear), 2012.
[21] D. Miln e and I. H. Witten, “An effective, low-cost meas-
ure of semantic relatedness obtained from wikipedia
links,” in Proceeding of AAAI Workshop on Wikipedia
and Artificial Intelligence: an Evolving Synergy, 2008, pp.
[22] D. Milne, “An open-source toolkit for mining Wikipe-
dia,” in Proceeding of New Zealand Computer Science
Research Student Conference, vol. 9, 2009.