AN EFFICIENT APPROACH FOR SEMANTICALLYENHANCED DOCUMENT CLUSTERING BY USING WIKIPEDIA LINK STRUCTURE

International Journal of Artificial Intelligence & Applications (IJAIA), Vol. 5, No. 6, November 2014
DOI : 10.5121/ijaia.2014.5605 53
AN EFFICIENT APPROACH FOR SEMANTICALLY-
ENHANCED DOCUMENT CLUSTERING BY USING
WIKIPEDIA LINK STRUCTURE
Iyad AlAgha1
and Rami Nafee2
1
Faculty of Information Technology, The Islamic University of Gaza, Palestine
2
Faculty of Information Technology, The Islamic University of Gaza, Palestine
ABSTRACT
Traditional techniques of document clustering do not consider the semantic relationships between words
when assigning documents to clusters. For instance, if two documents talking about the same topic do that
using different words (which may be synonyms or semantically associated), these techniques may assign
documents to different clusters. Previous research has approached this problem by enriching the document
representation with the background knowledge in an ontology. This paper presents a new approach to
enhance document clustering by exploiting the semantic knowledge contained in Wikipedia. We first map
terms within documents to their corresponding Wikipedia concepts. Then, similarity between each pair of
terms is calculated by using the Wikipedia's link structure. The document’s vector representation is then
adjusted so that terms that are semantically related gain more weight. Our approach differs from related
efforts in two aspects: first, unlink others who built their own methods of measuring similarity through the
Wikipedia categories; our approach uses a similarity measure that is modelled after the Normalized
Google Distance which is a well-known and low-cost method of measuring term similarity. Second, it is
more time efficient as it applies an algorithm for phrase extraction from documents prior to matching terms
with Wikipedia. Our approach was evaluated by being compared with different methods from the state of
the art on two different datasets. Empirical results showed that our approach improved the clustering
results as compared to other approaches.
KEYWORDS
Document Clustering, Semantic Similarity, Ontology, Wikipedia.
1. INTRODUCTION
Traditional techniques of document clustering are often based on word co-occurrence while it
ignores the semantic relationships between words[1,2]. Thus, if two documents use different
collections of words to represent the same topic, the clustering approach will consider these
documents different and will assign them to different clusters despite the semantic similarity
between the core words.
Exiting research on document clustering has approached the lack of semantics by integrating
ontologies as background knowledge [3,4]. An Ontology is a hierarchy of domain terms related
via specific relations. Ontologies can improve document clustering by identifying words that are
probably synonyms or semantically related even though they are syntactically different[5].
Comparing terms in documents using ontology relies on the fact that their corresponding concepts
within the ontology may have properties in the form of attributes, level of generality or
specificity, and their relationships.

54
WordNet is an example of ontologies that is widely used as background knowledge for document
clustering. Many efforts have explored the use of WordNet to incorporate semantics into the bag
of words (BOW)[6]. This can be done by matching and replacing words in the document with the
most appropriate terms from WordNet. Then, the attributes of and relationships between these
terms can be exploited to enhance clustering. However, WordNet-based similarity has not proven
to enhance clustering as first expected. While some works showed that WordNet has the potential
to improve clustering results[5,7,8], other works have reported that the ontological concepts add
no value and sometimes impair performance of document clustering[9]. Besides, some
researchers indicate that WordNet does not provide good word similarity data and that its
structure does not fit well for this task[10].
Some clustering techniques have used domain ontologies such as Mesh Ontology to cluster
documents related to specific domains of knowledge[3]. However, these ontologies have limited
coverage and it is unlikely to cover all concepts mentioned in the document collections, especially
when documents are from general domain.
Rather than relying on domain specific and limited coverage ontologies, some approaches have
used Wikipedia concepts and category information to enrich document representation and handle
the semantic relationships between document terms[11-13]. Wikipedia is much more
comprehensive than other ontologies since it capture a wide range of domains, is frequently
updated and well structured. Wikipedia can be seen as an ontology where each article represents
a single ontology concept, and all concepts are linked together by hyperlinks. In addition,
Wikipedia has a hierarchal categorization that resembles the structure of an ontology whereas
each article belongs to one or more information categories.
In this paper, we propose an approach to improve document clustering by explicitly incorporating
the semantic similarity between Wikipedia concepts into the document’s vector space model. Our
approach is distinguished over similar approaches in terms of the way we used to efficiently map
the document content to Wikipedia concepts and the low-cost measure we adapted to determine
semantic similarity between terms. In the following section, we discuss similar efforts that also
exploited knowledge from Wikipedia to enhance document clustering, and compare their
approaches with ours.
2. RELATED WORK
Techniques that employ Ontological features for clustering try to integrate the ontological
background knowledge into the clustering algorithm. Ontology based similarity measures are
often used in these techniques to calculate the semantic similarity between document terms. There
is a plenty of Ontology-based similarity measures that exploit different ontological features, such
as distance, information content and shared features, in order to quantify the mutual information
between terms (reader is referred to [3] for a review and comparison of ontology based similarity
measures). Distance between two document vectors is then computed based on the semantic
similarity of their terms.
A number of research efforts explored the use of Wikipedia to enhance text mining tasks,
including document clustering[11,14,15], text classification [16] and information retrieval[17].
Few approaches have explored utilizing Wikipedia as a knowledge base for document clustering.
Gabrilovich et al. [18] proposed and evaluated a method that is based on matching documents
with the most relevant articles of Wikipeda, and then augmenting the document’s BOW with the
semantic features. Spanakis et al. [19] proposed a method for conceptual hieratical clustering that
exploits Wikipedia textual content and link structure to create compact document representation.

55
However, these efforts do not make use of the structural relations in Wikipeida. As a result, the
semantic relatedness between words that are not synonyms is not considered when computing the
similarity between documents.
Huang et al. [13] proposed an approach that maps terms within documents to Wikipedia’s anchors
vocabulary. Then they incorporated the semantic relatedness between concepts by using Milne
and Witten measure [20] which takes into account all of the Wikipedi’s hyperlinks. Our work is
similar in that it also uses a similarity measure that is based on the Wikipedia’s hyperlinks.
However, their approach did not tackle the issue of frequent itemsets, and they instead used a less
efficient approach by examining all possible n-grams. Another difference is the way the
document similarity is measured: while they augmented the measure of document similarity, our
approach augments the document’s vector by reweighting the tf-idf score of each word according
to its relatedness to other document’s words. This makes our approach independent of, and can be
used with, any measure of document similarity since the reweighting process is carried out before
computing similarity between document pairs.
Another work that can be compared to ours is presented by Hu et al.[11]. They developed two
approaches: exact match and relatedness-match, to map documents to Wikipedia concepts, and
further to Wikipedia categories. Then documents are clustered based on a similarity metric which
combines document content information, concept information as well as category information.
However, their approach requires pre-processing of the whole Wikipedia’s textual content, a
thing that leads to substantial increase in both runtime and memory requirements. Instead, our
approach does not require any access to the Wikipedia’s textual content, and relies only on the
Wikipedia’s link structure to compute similarity between terms.
Hu et al. [12] proposed a method that mines synonym, hypernym and associative relations from
Wikipedia, and append that to traditional text similarity measure to facilitate document clustering.
However, their method was developed specifically for the task and has not been investigated
independently. They also built their own method of measuring similarity through Wikipedia’s
category links and redirects. We instead used a similarity measure that is modeled after the
Normalized Google Distance [21] which is a well-known and low-cost method of measuring
similarity between terms based on the link structure of the Web.
Wikipedia has been employed in some efforts for short text classification. For example, Hu et al.
[22] proposed an approach that generates queries from short text and use them to retrieve accurate
Wikipedia pages with the help of a search engine. Titles and links from the retrieved pages are
then extracted to serve as additional features for clustering. Phan et al. [23] presented a
framework for building classifiers that deal with short text. They sought to expand the coverage
of classifiers by topics coming for external knowledge base (e.g. Wikipedia) that do not exist in
small training datasets. These approaches, however, use Wikipedia concepts without considering
the hierarchical relationships and categories embedded in Wikipedia.
3. AN APPROACH FOR WIKIPEDIA-BASED DOCUMENT CLUSTERING
The pseudo code of our approach for Wikipedia-based document clustering is shown in Figure 1,
and consists of three phases: The first phase includes of a set of text processing steps for the
purpose of determining terms that best represent the document content. In the second phase, each
document is represented by using the tf-idf weighted vector. Document terms are then mapped to
Wikipedia concepts. In the third phase, the similarity between each pair of Wikipedia concepts is
measured by using the Wikipedia link structure. The tf-idf weights of original terms are then
reweighted to incorporate the similarity scores obtained from Wikipedia. By the end of the
algorithm, the tf-idf representation of each document is enriched so that terms that are

56
Input: A set of documents
Begin
{Phase 1: Pre-processing and extraction of frequent phrases}
for each document d D do
Apply stemming and stopword removal
end for
Concatenate all documents
Apply Apriori algorithm to extract frequent itemsets
{Phase 2: Construct tf-idf weighted vectors and map terms to Wikipedia concepts}
Tokenize d
Discard tokens that overlap with frequent phrases
Discard rare terms
Build the BOW of d where BOW = Retained tokens ⋃ frequent phrases
for each term t BOW do
Calculate tf-idf for t
if t matches Wikipedia concept(s) then
Replace t with matching Wikipedia concept(s)
end if
end for
end for
{Phase 3: {Reweighting tf-idf weights}
for each term BOW of d do
for each term BOW of d AND do
Compute similarity between using equation 1
if are ambiguous then
Perform word-sense disambiguation (see section 6)
end if
end for
Reweight using equation 2
end for
end for
{Document clustering}
Apply any conventional clustering algorithm (e.g. k-means)
End
Figure 1. Pseudo code of our algorithm of document clustering
semantically related gain more weight. Documents can then be clustered using any traditional
clustering method such as k-means. These phases are explained in detail in the subsequent
sections.
Prior to applying our approach, Wikipedia’s vocabulary of anchor text is retrieved from the
Wikipedia dump, which is a copy of all Wikipedia content, and stemmed in order to be
comparable with the stemmed document content. For measuring similarity between Wikipedia
concepts, all outgoing hyperlinks, incoming hyperlinks and categories of articles are also
retrieved. Note that this task incurs a one-time cost, thus allowing the clustering algorithm to be
invoked multiple times without the additional overhead of reprocessing the Wikipedia content.

57
4. CONSTRUCTION OF DOCUMENT’S VECTOR SPACE MODEL
The first step of our clustering approach is to represent each document as a bag of words (BOW).
Note that traditional clustering algorithms treat a document as a set of single words, thus losing
valuable information about the meaning of terms. When incorporating semantics in document
clustering, it is necessary to preserve phrases, the consecutive words that stand together as a
conceptual unit. Without preserving phrases, actual meanings of document terms may be lost,
making it difficult to measure semantic similarity in between. For example, the phrase “big bang
theory” refers to a concept that is entirely different from what its individual tokens refer to. Thus,
we aim to create a document’s BOW representation whose attributes include not only single
words but also phrases that have standalone meanings. This phase starts with some standard text
processing operations including stopword removal and word stemming. Stopwords are words that
occur frequently in documents and have little informational meanings. Stemming finds the root
form of a word by removing its suffix. In the context of mapping with Wikipedia concepts,
stemming allows to recognize and deal with variations of the same word as if they were the same,
hence detecting mappings between words with the same stem.
To construct the document’s bag of words and phrases, we used a simple method based on
Apriori algorithm [24] to find frequent occurring phrases from a document collection. A phrase is
defined as frequent if it appears in n number of documents (For our task we set n = 3). The
Apriori algorithm consists of two steps: In the first step, it extracts frequent itemsets, or phrases,
from a set of transactions that satisfy a user-specified minimum support. In the second step, it
generates rules from the discovered frequent itemsets. For this task, we only need the first step,
i.e., finding frequent itemsets. In addition, the algorithm was restricted to find itemsets with four
words or fewer as we believe that most Wikipedia concepts contain no more than four words (this
restriction can be easily relaxed).
After extracting frequent itemsets, we perform word tokenization to break the document text into
single words. Many of the resulting tokens can be already part of the extracted itemsets.
Therefore, we remove tokens that overlap with any of the retrieved frequent itemsets. Stemmed
tokens as well as frequent itemsets that occur in the document will be combined together to form
the BOW representing the document. Rare terms that infrequently occur in the document
collection can introduce noise and degrade performance. Thus, terms that occur in the document
collection less than or equal to a predefined threshold are discarded from the document’s BOW.
It is worth noting here that similar works that exploited Wikipedia for document clustering often
did not consider mining frequent itemsets occurring in the document [11 , 13 ]. Instead, they
extract all possible n-grams from the document by using a sliding widow approach and match
them with the Wikipedia content. In contrast, our approach of extracting frequent itemset prior to
the concept-mapping process is more time-efficient as it avoids the bottleneck of matching all
possible n-grams to Wikipedia concepts.
After identifying the document’s BOW, the next step is to map terms within the BOW to
Wikipedia concepts: Each term is compared with Wikipedia anchors, and matching terms are
replaced by the corresponding Wikipedia concepts. Terms that do not match any Wikipedia
concept are not discarded from the BOW in order to avoid any noise or information loss.
Formally, let be a set of documents and be the set of different
terms occurring in a document . Note that T includes both: 1) Wikipedia concepts which replace
original terms after the mapping process. 2) Terms that do not match any Wikipedia concepts.
The weight of each document term is then calculated using tf-idf (term frequency-inverted
document frequency). Tf-idf weighs the frequency of a term in a document with a factor that

58
discounts its importance when it appears in almost all documents. The tf-idf of term in
document is calculated using the following equation:
The document’s vector representation is then constructed from the tf-idf weights of its terms:
5. MEASURING SEMANTIC SIMILARITY BETWEEN WIKIPEDIA TERMS
After representing the document as a vector of term tf-idf weights, the next step is to augment
these weights so that terms gain more importance according to their semantic similarity to the
other document terms.
To measure similarity between document terms, we used a measure that is based on the
Normalized Google Distance Measure (NGD) [21]. The NGD measure first uses the Google
search engine to obtain all Web pages mentioning these terms. Pages that mention both terms
indicate relatedness, while pages that mention only one term indicate the opposite. The NGD
denotes the distance or dissimilarity between two terms: the smaller the value of NGD, the more
related the terms are semantically. For this work, the measure is adapted to exploit Wikipedia
articles instead of the Google’s search results. Formally, the Wikipedia-based similarity measure
is:
(1)
where s and t are a pair of Wikipedia concepts. S and T are the sets of all Wikipedia articles that
link to s and t respectively, and R is set of all Wikipedia concepts. The output of this measure
ranges between 0 and 1, where values close to 1 denote related terms while values close to 0
denote the opposite. Note that the advantage of this measure is its low computational cost since it
only considers the links between Wikipedia articles to define similarity.
After computing the similarity between each pair of terms, the tf-idf weight of each term is
adjusted to consider its relatedness to other terms within the document’s vector representation.
The adjusted weight of a term is calculated using the following equation:
(2)
where is the semantic similarity between the terms and , and is calculated using
Equation 1. N is the number of co-occurred terms in document d. The threshold denotes the
minimum similarity score between two terms. Since we are interested in emphasizing more
weight on terms that are more semantically related, it is necessary to set up a threshold value.
Note that this measure assigns an additional weight to the original term’s weight based on its
similarity to other terms in the document. The term weight remains unchanged if it is not related
to any other term in the document or if it is not mapped to any Wikipedia concept. The final
document’s vector is:

59
After constructing the semantically-augmented vectors for all documents, any conventional
measure of document similarity, such as the cosine measure, can be used to measure similarity
between document pairs. Note that in our approach we incorporate the similarity scores in the
document representation before applying the document similarity measure. Thus, our approach is
independent of, and hence can be used with, any similarity measure or clustering algorithm.
6. WORD SENSE DISAMBIGUATION
Concepts mentioned in Wikipedia are explicitly linked to their corresponding articles through
anchors. These anchors can be considered as sense annotations for Wikipedia concepts.
Ambiguous words such as “eclipse” are linked to different Wikipedia articles based on their
meanings in the context where they occur (e.g. eclipse "astronomical event", eclipse "software
suite", eclipse "foundation"). When mapping document terms to Wikipedia concepts, it is
necessary to perform word sense disambiguation to identify the correct word sense. Failing to do
so may result in false results when measuring similarity between terms.
One way to disambiguate words is to simply use the most common sense. The commonness of a
sense is identified by the number of anchors that link to it in Wikipedia. For example, over 95%
of anchors labelled as “Paris” link to the capital of France while the rest link to other places,
people or even music. However, choosing the most common sense is not enough and it is not
always the best decision. Instead, we used the same approach used in [21] which uses the two
terms involved in the similarity measure to disambiguate each other. This is done by selecting the
two candidate senses that most closely related to each other. We start by choosing the top
common senses for each term (For simplicity, we ignore senses that contribute with less than 1%
of the anchor’s links). We then measure the similarity between every pair of senses, and the two
senses with the highest similarity score are considered.
7. EVALUATION
Wikipedia releases its database dumps periodically, which can be downloaded from
http://download.wikipedia.org. The Wikipedia dump used in this evaluation was released on the
13th August 2014, and contains 12100939 articles. The data was presented in XML format. We
used the WikipediaMiner [25] toolkit to process the data and extract the categories and outlinks
out of Wikipedia dump.
7.1. Methodology
Our objective was to compare our approach with other approaches from the state of the art.
Therefore, we used the same evaluation settings used by Hu et al. [12] in order to make our
results comparable with theirs. The following two test sets were created:
• Reuters-21578 contains short news articles. The subset created consists of categories in the
original Reuters dataset that have at least 20 and at most 200 documents. This results in 1658
documents and 30 categories in total.
• OHSUMed contains 23 categories and 18302 documents. Each document is the concatenation
of title and abstract of a medical science paper.
Besides our method, we implemented and tested three different text representation methods, as
defined below:
• Bag of Words: The traditional BOW method with no semantics. This is the baseline case.

60
• Hotho et al.’s method: this is a reimplementation of Hotho et al.’s WordNet-based
algorithm[8]. The intention of considering this method is to compare how the use of Wikipedia
as background knowledge influences the clustering results as compared to WordNet.
• Hu et al.: this is a reimplementation of the Hu et al.’s algorithm [12] which leverages
Wikipedia as a background knowledge for document clustering. While Hu et al. built their
own measure of document similarity based on Wikipedia’s links, our approach uses a
similarity measure that is based on the Normalized Google Distance.
To focus our investigation on the representation rather than the clustering method, the standard k-
means clustering algorithm was used. We used two evaluation metrics: Purity and F-score. Purity
assumes that all samples of a cluster are predicted to be members of the actual dominant class for
that cluster. F-score combines the information of precision and recall which is extensively applied
in information retrieval.
7.2. Results
Table 1 shows how the different methods perform in clustering on the two datasets. In general,
the performance of BOW on both datasets is improved by incorporating background knowledge
either from WordNet (Hotho et al.’s method) or Wikipedia (Hu et al. and our method). For
instance, according to the F-score, for the Reuters dataset, our method and Hotho et al’s method
achieve 27.23% and 14.71% respectively.
On comparing the use of Wikipedia to WordNet, our approach and Hu et al.’s, which both use
Wikipedia to enrich document similarity, outperformed the Hotho et al.’s approach for both
datasets. This demonstrates the potential of integrating Wikipedia as a knowledge source as
compared to the WordNet based method.
Comparing our approach with the other three methods, our approach achieves the best F-score
and purity on both datasets. We applied t-test to compare between the performance of our
approach and the others. Results show that our approach significantly outperformed all other
methods on the Reuters dataset with the p-value < 0.05. When using the OHSUMed dataset, the
difference between our approach and the BOW and Hotho et al. was also significant. However,
there was no significant difference between the Hu et al. approach and ours (p < 0.32). In general,
our approach provides slight improvement over the Hu et al.’s method. As both approaches
exploit the Wikipedia link structure to enrich the document representation, we believe that the
slight improvement in our approach stems from the adopted similarity measure that is based on
the well-known Normalized Google Distance.
Table 1. Comparison with related work in terms of purity and F-score
Reuters 21578 OHSUMed
Purity (Impr.) F-score (Impr.) Purity (Impr.) F-score (Impr.)
Bag of Words 0.571 0.639 0.36 0.470
Hotho et al. 0.594 (4.02%) 0.685 (7.20%) 0.405 (12.5%) 0.521 (10.85%)
Hu et al. 0.655 (8.40%) 0.733 (14.71%) 0.475 (31.94%) 0.566 (20.43%)
Our
Approach
0.703 (23.11%) 0.813 (27.23%) 0.483 (34.16%) 0.592 (25.95%)

61
8. CONCLUSION AND FUTURE WORK
In this work, we proposed an approach for leveraging Wikipedia link structure to improve text
clustering performance. Our approach uses a phrase extraction technique to efficiently map
document terms to Wikipedia concepts. Afterwards, the semantic similarity between Wikipedia
terms is measured by using a measure that is based on the Normalized Google Distance and the
Wikipedia’s link structure. The document representation is then adjusted so that each term is
assigned an additional weight based on its similarity to other terms in the document. Our
approach differs from similar efforts from the state of the art in two aspects: first, unlink other
works that built their own methods of measuring similarity through the Wikipedia’s category
links and redirects, we instead used a similarity measure that is modelled after the Normalized
Google Distance which is a well-known and low-cost method of measuring similarity between
terms. Second, while other approaches used to match all possible n-grams to Wikipedia concepts,
our approach is more time efficient as it applies an algorithm for phrase extraction from
documents prior to matching terms with Wikipedia. In addition, our approach does not require
any access to the Wikipedia’s textual content, and relies only on the Wikipedia’s link structure to
compute similarity between terms. The proposed approach was evaluated by being compared with
different methods from related work (e.g. Bag of Words with no semantics, clustering with
WordNet as well as clustering with Wikipedia) on two datasets: Reuters 21578 and OHSUMed.
We think that our approach can be extended to other applications that are based on the
measurement of document similarity, such as information retrieval and short text classification. In
our future work, we aim to further improve our concept-mapping technique: Currently, only the
document terms that exactly match Wikipedia concepts are extracted and used for the similarity
measure. Instead of exact matching, we aim to utilize the graph of Wikipedia links to build the
connection between Wikipedia concepts and the document content even if they cannot exactly
match. This approach can be more useful when Wikipedia concepts cannot fully cover the
document content.
REFERENCES
[1] Hotho, Andreas, Maedche, Alexander & Staab, Steffen, (2002) "Ontology-Based Text Document
Clustering", KI, Vol. 16, No. 4: pp. 48-54.
[2] Charola, Apeksha & Machchhar, Sahista, (2013) "Comparative Study on Ontology Based Text
Documents Clustering Techniques", Data Mining and Knowledge Engineering, Vol. 5, No. 12: pp.
426.
[3] Zhang, Xiaodan, Jing, Liping, Hu, Xiaohua, Ng, Michael & Zhou, Xiaohua, (2007) "A Comparative
Study of Ontology Based Term Similarity Measures on Pubmed Document Clustering", Advances in
Databases: Concepts, Systems and ApplicationsSpringer. pp. 115-126.
[4] Logeswari, S & Premalatha, K, (2013) "Biomedical Document Clustering Using Ontology Based
Concept Weight", Computer Communication and Informatics (ICCCI), 2013 International Conference
on: IEEE. pp. 1-4.
[5] Hotho, Andreas, Staab, Steffen & Stumme, Gerd, (2003) "Ontologies Improve Text Document
Clustering", Data Mining, 2003. ICDM 2003. Third IEEE International Conference on: IEEE. pp.
541-544.
[6] Wei, Tingting, Zhou, Qiang, Chang, Huiyou, Lu, Yonghe & Bao, Xianyu, (2014) "A Semantic
Approach for Text Clustering Using Wordnet and Lexical Chains", Expert Systems with
Applications, Vol. 42, No. 4,
[7] Wang, Pu, Hu, Jian, Zeng, Hua-Jun, Chen, Lijun & Chen, Zheng, (2007) "Improving Text
Classification by Using Encyclopedia Knowledge", Data Mining, 2007. ICDM 2007. Seventh IEEE
International Conference on: IEEE. pp. 332-341.
[8] Hotho, Andreas, Staab, Steffen & Stumme, Gerd, (2003) "Wordnet Improves Text Document
Clustering", Proceedings of SIGIR Semantic Web Workshop: ACM. pp. 541-544.
[9] Moravec, Pavel, Kolovrat, Michal & Snasel, Vaclav, (2004) "Lsi Vs. Wordnet Ontology in
Dimension Reduction for Information Retrieval", Dateso. pp. 18-26.

62
[10] Passos, Alexandre & Wainer, Jacques, (2009) "Wordnet-Based Metrics Do Not Seem to Help
Document Clustering", Proc. of the of the II Workshop on Web and Text Intelligente, S‫£أ‬o Carlos,
Brazil. pp.
[11] Hu, Xiaohua, Zhang, Xiaodan, Lu, Caimei, Park, Eun K & Zhou, Xiaohua, (2009) "Exploiting
Wikipedia as External Knowledge for Document Clustering", Proceedings of the 15th ACM SIGKDD
international conference on Knowledge discovery and data mining: ACM. pp. 389-396.
[12] Hu, Jian, Fang, Lujun, Cao, Yang, Zeng, Hua-Jun, Li, Hua, Yang, Qiang & Chen, Zheng, (2008)
"Enhancing Text Clustering by Leveraging Wikipedia Semantics", Proceedings of the 31st annual
international ACM SIGIR conference on Research and development in information retrieval: ACM.
pp. 179-186.
[13] Huang, Anna, Milne, David, Frank, Eibe & Witten, Ian H, (2009) "Clustering Documents Using a
Wikipedia-Based Concept Representation", Advances in Knowledge Discovery and Data
MiningSpringer. pp. 628-636.
[14] Kumar, Kiran, Santosh, GSK & Varma, Vasudeva, (2011) "Multilingual Document Clustering Using
Wikipedia as External Knowledge": Springer
[15] Roul, Rajendra Kumar, Devanand, Omanwar Rohit & Sahay, SK, (2014) "Web Document Clustering
and Ranking Using Tf-Idf Based Apriori Approach", arXiv preprint arXiv:1406.5617, Vol., No.
[16] Zhang, Zhilin, Lin, Huaizhong, Li, Pengfei, Wang, Huazhong & Lu, Dongming, (2013) "Improving
Semi-Supervised Text Classification by Using Wikipedia Knowledge", Web-Age Information
ManagementSpringer. pp. 25-36.
[17] Han, May Sabai, (2013) "Semantic Information Retrieval Based on Wikipedia Taxonomy",
International Journal of Computer Applications Technology and Research, Vol. 2, No. 1: pp. 77-80.
[18] Gabrilovich, Evgeniy & Markovitch, Shaul, (2006) "Overcoming the Brittleness Bottleneck Using
Wikipedia: Enhancing Text Categorization with Encyclopedic Knowledge", AAAI. pp. 1301-1306.
[19] Spanakis, Gerasimos, Siolas, Georgios & Stafylopatis, Andreas, (2012) "Exploiting Wikipedia
Knowledge for Conceptual Hierarchical Clustering of Documents", The Computer Journal, Vol. 55,
No. 3: pp. 299-312.
[20] Milne, David & Witten, Ian H, (2008) "Learning to Link with Wikipedia", Proceedings of the 17th
ACM conference on Information and knowledge management: ACM. pp. 509-518.
[21] Cilibrasi, Rudi L & Vitanyi, Paul MB, (2007) "The Google Similarity Distance", Knowledge and
Data Engineering, IEEE Transactions on, Vol. 19, No. 3: pp. 370-383.
[22] Hu, Xia, Sun, Nan, Zhang, Chao & Chua, Tat-Seng, (2009) "Exploiting Internal and External
Semantics for the Clustering of Short Texts Using World Knowledge", Proceedings of the 18th ACM
conference on Information and knowledge management: ACM. pp. 919-928.
[23] Phan, Xuan-Hieu, Nguyen, Le-Minh & Horiguchi, Susumu, (2008) "Learning to Classify Short and
Sparse Text & Web with Hidden Topics from Large-Scale Data Collections", Proceedings of the 17th
international conference on World Wide Web: ACM. pp. 91-100.
[24] Agrawal, Rakesh & Srikant, Ramakrishnan, (1994) "Fast Algorithms for Mining Association Rules",
Proc. 20th int. conf. very large data bases, VLDB. pp. 487-499.
[25] "Wikipedia Miner". 20th September 2014]; Available from: http://wikipedia-
miner.cms.waikato.ac.nz/.
Authors
Iyad M. AlAgha received his MSc and PhD in Computer Science from the University
of Durham, the UK. He worked as a research associate in the center of technology
enhanced learning at the University of Durham, investigating the use of Multi-touch
devices for learning and teaching. He is currently working as an assistant professor at
the Faculty of Information technology at the Islamic University of Gaza, Palestine. His
research interests are Semantic Web technology, Adaptive Hypermedia, Human-
Computer Interaction and Technology Enhanced Learning.
Rami H. Nafee received his BSc from Al-Azhar University-Gaza and his MSc degree
in Information technology from the Islamic University of Gaza. He works as a Web
programmer at Information Technology Unit at Al-Azhar University-Gaza. He is also a
lecturer at the Faculty of Intermediate Studies at Al-Azhar University. His research
interests include Data Mining and Semantic Web.

AN EFFICIENT APPROACH FOR SEMANTICALLYENHANCED DOCUMENT CLUSTERING BY USING WIKIPEDIA LINK STRUCTURE

More Related Content

What's hot (10)

Viewers also liked (13)

Similar to AN EFFICIENT APPROACH FOR SEMANTICALLYENHANCED DOCUMENT CLUSTERING BY USING WIKIPEDIA LINK STRUCTURE (20)

Recently uploaded (20)

AN EFFICIENT APPROACH FOR SEMANTICALLYENHANCED DOCUMENT CLUSTERING BY USING WIKIPEDIA LINK STRUCTURE