Context-based Disambiguation using Wikipedia
Hugo Batista
1
, David Carrega
1
, Rui Rodrigues
1
and Joaquim Filipe
2
1
INSTICC, Set
´
ubal, Portugal
2
Escola Superior de Tecnologia, Instituto Polit
´
ecnico de Set
´
ubal, Set
´
ubal, Portugal
Keywords:
Ontology, Ambiguity, Context-awareness, Semantic Relatedness.
Abstract:
This paper addresses the problem of semantic ambiguity, identified in a previous work where we presented
an algorithm for quantifying semantic relatedness of entities characterized by a set of features, potentially
ambiguous. We propose to solve the feature ambiguity problem by determining the context defined by the
non-ambiguous features and then using this context to select the most adequate interpretation of the ambiguous
features. As a result, the entity semantic relatedness process will be improved by reducing the probability of
using erroneous features due to ambiguous meaning.
1 INTRODUCTION
In a previous work, we proposed a semantic relat-
edness measure between scientific concepts, using
Wikipedia
1
categories network as an ontology, based
on the length of the category path (Medina et al.,
2012).
Semantic relatedness between two concepts indi-
cates the degree in which these concepts are related
in a conceptual network, by computing not only their
semantic similarity but actually any possible seman-
tic relationship between them (Ponzetto and Strube,
2007) (Gracia and Mena, 2008).
The proposed measure considers not merely the
number of arcs in the graph between the nodes that
represent each concept, but also their relationship in
the taxonomy. This procedure has been extended to
measure semantic relatedness between entities, an en-
tity being defined as a set of features, i.e. concepts.
After observing a substantial number of features
were mapped to a disambiguation pages, it was con-
cluded that if we manage to discover the right feature
page mapping, the overall quality of results will im-
prove.
The disambiguation process in Wikipedia intends
to resolve the conflicts that arise when a single term
refers to more than one subject covered by Wikipedia.
For example, the word ”Matrix” can refer to a mathe-
matics topic, a movie, music albums, and many other
things.
1
http://en.wikipedia.org
Now we attempt to improve this process by adding
the feature of automatic page disambiguation.
The remaining sections of this document are or-
ganized as follows: in Section 2 we describe related
work in this area; Section 3 presents the problem
that derived the proposed solution; Section 4 presents
the proposed method of disambiguation of Wikipedia
pages based on context and in section 5 are presented
the results obtained after applying this method to an
entity with several features. Finally, in Section 6 we
draw the main conclusions and identify opportunities
for future work.
2 RELATED WORK
Semantic relatedness measures in hierarchical tax-
onomies can be categorized into three types (Slimani
et al., 2006):
1. Information Content or Node-based: evaluation of
the information content of a concept represented
by a node such as described in (Resnik, 1999).
The semantic relatedness between two concepts
reflects the amount of shared information between
them, generally in the form of their least common
subsumer (LCS).
2. Path or Edge-based: evaluation of the distance
that separates concepts by measuring the length
of the edge-path between them (Wu and Palmer,
1994) (Rada et al., 1989). A weight is assigned to
each edge, being that the calculated weight must
549
Batista H., Carrega D., Rodrigues R. and Filipe J..
Context-based Disambiguation using Wikipedia.
DOI: 10.5220/0005187605490553
In Proceedings of the International Conference on Knowledge Discovery and Information Retrieval (SSTM-2014), pages 549-553
ISBN: 978-989-758-048-2
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
reflect some of the graph properties (network den-
sity, node depth, link strength, etc.) (Jiang and
Conrath, 1997).
3. Hybrid: a combination of the former two (Jiang
and Conrath, 1997) (Leacock and Chodorow,
1998).
Regarding the ambiguity, we have a couple of
works that tried to solve this problem of human lan-
guages by using Wikipedia. Word sense ambiguity
exists in all natural languages across the world. One
of the first approaches uses Wikipedia to compare lex-
ical context around the ambiguous concept to the can-
didates of desambiguation at Wikipedia (Bunescu and
Pasca, 2006).
Some authors explored the possibility of using
Wikipedia labels, definition on the disambiguation
pages and Wordnet definitions combined to learn the
real true meaning of the sentences (Mihalcea, 2007).
Lexical databases, such as WordNet, have been
explored as knowledge bases to measure the semantic
similarity between words or expressions. However,
WordNet provides generic definitions and a somewhat
rigid categorization that does not reflect the intuitive
semantic meaning that a human might assign to a con-
cept.
Other works in this particular field aim to combine
the traditional approaches with the Wikipedia infor-
mations as an auxiliary source, to improve the results
(Ratinov et al., 2011). One of most common problems
with this kind of approaches rely on the time that it is
needed to perform the calculation. With that in mind,
the tests were reduced to a limited set of Wikipedia
information.
Based on that information we believe that a great
progress on disambiguation problem using Wikipedia
as base is still achievable.
3 PROBLEM
Currently Wikipedia is mainly used has a tool to ex-
tract semantic knowledge, having currently over 4
million articles and a well structured category net-
work, which allows us to extract the necessary infor-
mation to disambiguate an ambiguous term.
In our particular case, we have a generic entity,
this entity contains a list of features that describe her.
We want to find a Wikipedia article that represents
the semantic content of each feature. The problem is
that some topics lead us to disambiguation page
2
, a
non-article page which lists the various meanings the
2
http://en.wikipedia.org/wiki/Wikipedia:Disambiguation
ambiguous term and links to the articles which cover
them.
The challenge is to find the most appropriate arti-
cle from the list provided by the disambiguation page
with an acceptable time and efficiency.
To disambiguate an article it is first necessary a
context, the context will consist of all non-ambiguous
articles in the list of features, this context will be used
to calculate the proximity between him and every dis-
ambiguation article, the article closest context should
be the most suitable article.
In short, our problem is to find Wikipedia articles
that semantically represent the features and disam-
biguate the ambiguous articles quickly and efficiently.
4 PROPOSED DISAMBIGUATION
METHOD
Search for Articles
Considering the problem described in the previous
section, it is first necessary to find an article that se-
mantically represent each feature. There are two basic
ways to find articles from a feature:
1. Find an Wikipedia article directly from the feature
literally comparing the text of the feature with the
title of the article.
2. Decomposing the feature, in order to obtain
simpler sub-features and use them to find the
Wikipedia articles, this technique can lead to se-
mantic deviations, so it should be avoided or care-
fully treated.
The solution we found was to develop a set of
methods that can meet efficiency about 60% or higher
of the article for the features.
The developed methods are:
• Direct Search: Find an Wikipedia article, liter-
ally comparing the text of the feature, singular and
plural, with the title of the article. This method is
reused by the other methods.
• Regex
3
: Find regular expressions in the text of
the feature, and treats it according to the type of
regular expressions found. Much of the Regex
are developed to decompose the features contain-
ing in it’s text the word ”and” or punctuation
mark’s like ”comma” or ”colon”; These elements
are very common and easily decomposed because
they generate predictable structures.
3
http://en.wikipedia.org/wiki/Regular expression
KDIR2014-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval
550
This technique can be considered a decomposi-
tion, but has a small probability of carrying a se-
mantic deviation, taking into account that there is
no loss or change of words.
• Decomposition: This method decomposes the
features making successive sweeps, in which the
number of words considered is equal to the num-
ber of words in the text of the feature least one
unit by sweeping, until find a sub-feature since it
has at least two words.
This method has lead to semantic deviations since
there is loss of words which may partially or to-
tally change the semantics. This effect is mini-
mized by maximizing the number of words to con-
sider in each sweep.
Disambiguate
After searching for articles for each unambiguous fea-
ture, considering that it was found at least one ar-
ticle, we can now disambiguate the ambiguous fea-
tures. Otherwise it is not possible to disambiguate,the
context is a mandatory element in the act of disam-
biguation.
First it is necessary to know the Article Context
of each disambiguation article and the General Con-
text, Article Context is the context of just one article,
while the General Context is the context of all found
articles.
To set the context of an article it was decided to
use the categories of the article itself, one category
define a topic and a set of them can pin down a con-
text. The problem of using only their categories is that
it is not possible to know the categories with more or
less relevance to the article. To solve this problem was
considered joining the categories of article categories
of its outlinks.
• Article Context: is the count of repetitions of
each element of the union between its categories
and the categories of their outlinks. This approach
is a mix of those found in the papers (Milne and
Witten, 2008) and (Radhakrishnan and Varma,
2013). The Figure 1 illustrates the earlier descrip-
tion.
• General Context: is the count of repetitions of
each element of the union of the Article Context
of all found articles.
To get the best article for an ambiguous feature it’s
calculated the similarity between the Article Context
of each disambiguation article and the General Con-
text, using the similarity measure Cosine Similarity.
Figure 1: Illustration how to get the context of an article.
cos(Θ) =
A · B
k
A
kk
B
k
=
∑
n
i=1
A
i
B
i
q
∑
n
i=1
A
2
i
q
∑
n
i=1
B
2
i
(1)
The article with the highest similarity will be the
most suitable to disambiguate the feature. This pro-
cess is described in Figure 2.
5 RESULTS AND DISCUSSION
Results
The data presented in the next three pie charts are
drawn from a battery of tests consisting of 15 entities
with an average of 62 features, 3 of them ambiguous.
3%
39%
34%
24%
Disambiguation
Direct
Regex
Decomposition
The chart above shows the percentage of features
found with each type of process.
We can observe that most of the features are han-
dled by the Direct and Regex Processes, what is de-
sirable, as these processes have a small probability of
obtaining semantic deviations.
33%
67%
Not Found
Found
Context-basedDisambiguationusingWikipedia
551
Figure 2: Disambiguation based on context.
The chart above shows the percentage of features
found and not found.
Considering the demand for articles only directly,
the proposed solution shows a 42% improvement.
42%
58%
Not Disambiguated
Disambiguated
The chart above shows the percentage of disam-
biguated features and not disambiguated feature.
These results can be further improved by using
other techniques that have not yet been explored as
Clustering or Leftess (Gyllstrom and Moens, 2011),
these techniques can improve both effectiveness and
efficiency.
Practical Example
The Table 1 shows a set of features and the results
obtained from the developed algorithm.
Table 1 shows two ambiguous features the Tables
2 and 3 shows their disambiguation articles and the
similarity value between them and the General Con-
text.
We can observe from the data obtained that the
articles with a nearest context of the general context
have a value greater proximity unlike those containing
one farthest context.
In this particular case the features of the entity is
mostly under the topic database. The features ”Index”
and ”Table” should aim to articles strongly linked to
Table 1: List of features and their results.
Features Type of Process Results
Datamining Direct Datamining
Data Analytics Direct Data Analytics
Query Processing and Optimization Regex Query Processing
Query Optimization
Semi-structured and Unstructured Data Regex Semi-structured Data
Unstructured Data
WWW and Databases Regex WWW (World Wide Web)
Databases
Statistics Exploratory Data Analysis Decomposition Exploratory Data Analysis
Object-Oriented Database Systems Decomposition Object-Oriented Database
Database Systems
Large Scale Databases NO RESULT
NoSQL Databases NO RESULT
Table Ambiguous Table (database)
Index Ambiguous Database index
Table 2: Ambiguous feature ”Table” similarity values.
Disambiguation Page Similarity Value
Table (database) 0.135
Table (information) 0.070
Tables (board game) 0.021
Table (furniture) 0.010
Table (parliamentary procedure) 0.005
the database topic which is the case for the articles
”Table (database)” and ”Database index”.
6 CONCLUSIONS AND FUTURE
WORK
In this paper we used the Wikipedia Category Net-
work (WCN) with the link structure available to com-
pute the semantic relatedness of multiple meanings of
an ambiguous page trying to find the best possible ar-
ticle with the less time possible.
KDIR2014-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval
552
Table 3: Ambiguous feature ”Index” similarity values.
Disambiguation Page Similarity Value
Database index 0.132
Array data structure 0.024
Bibliographic index 0.016
Stock market index 0.005
Thumb index 0.000
This work allied with the previous work allows
the building off NLP applications that compare the se-
mantic relatedness of two generic entities in a viable
time with some degree of precision.
The results have shown a promising future al-
though we still need to test the results to human judg-
ment so that we can verify the veracity of our conclu-
sions.
Our proposal is based on the pre-processing of the
entire WCN, but in need of future work like:
1. Relevance of a category in a context (be it a arti-
cle or several articles), based on the Leftness as
described in (Gyllstrom and Moens, 2011).
2. Inclusion of inlinks and not only outlinks as re-
ferred in the paper, to increase the precision of the
article in the context.
3. Applying K-NN
4
Algorithm to create basic clus-
ters of articles based on the semantic relatedness
of categories between them. This will allow us to
avoid noise in the disambiguation process and to
also find outliers.
ACKNOWLEDGEMENTS
We gratefully acknowledge the support of the Poly-
technic Institute of Set
´
ubal and the School of Tech-
nology of Set
´
ubal.
REFERENCES
Bunescu, R. C. and Pasca, M. (2006). Using encyclope-
dic knowledge for named entity disambiguation. In
EACL’06, pages 9–16.
Gracia, J. and Mena, E. (2008). Web-based measure of se-
mantic relatedness. In WISE’08, pages 136–150.
Gyllstrom, K. and Moens, M.-F. (2011). Examining
the ’leftness’ property of wikipedia categories. In
CIKM’11, pages 2309–2312.
Jiang, J. J. and Conrath, D. W. (1997). Semantic similar-
ity based on corpus statistics and lexical taxonomy.
CoRR.
4
http://en.wikipedia.org/wiki/K-nearest neighbors algorithm
Leacock, C. and Chodorow, M. (1998). Combining local
context and wordnet similarity for word sense identi-
fication. In MIT Press, pages 265–283.
Medina, L. A. S., Fred, A. L. N., Rodrigues, R., and Filipe,
J. (2012). Measuring entity semantic relatedness using
wikipedia. In KDIR’12, pages 431–437.
Mihalcea, R. (2007). Using wikipedia for automatic word
sense disambiguation. In HLT-NAACL’07, pages 196–
203.
Milne, D. and Witten, I. H. (2008). An effective, low-
cost measure of semantic relatedness obtained from
wikipedia links. In In Proceedings of AAAI 2008.
Ponzetto, S. P. and Strube, M. (2007). Knowledge derived
from wikipedia for computing semantic relatedness. J.
Artif. Intell. Res. (JAIR), pages 181–212.
Rada, R., Mili, H., Bicknell, E., and Blettner, M. (1989).
Development and application of a metric on semantic
nets. IEEE Transactions on Systems, Man, and Cyber-
netics, pages 17–30.
Radhakrishnan, P. and Varma, V. (2013). Extracting seman-
tic knowledge from wikipedia category names. In Pro-
ceedings of the 2013 Workshop on Automated Knowl-
edge Base Construction, AKBC ’13, pages 109–114.
Ratinov, L.-A., Roth, D., Downey, D., and Anderson, M.
(2011). Local and global algorithms for disambigua-
tion to wikipedia. In ACL’11, pages 1375–1384.
Resnik, P. (1999). Semantic similarity in a taxonomy:
An information-based measure and its application to
problems of ambiguity in natural language. J. Artif.
Intell. Res. (JAIR), pages 95–130.
Slimani, T., Yaghlane, B. B., and Mellouli, K. (2006). A
new similarity measure based on edge counting. In
Proceedings of world academy of science, engineer-
ing and technology, volume 17.
Wu, Z. and Palmer, M. S. (1994). Verb semantics and lexi-
cal selection. In ACL’94, pages 133–138.
Context-basedDisambiguationusingWikipedia
553
