Arabic Text Classification using Bag-of-Concepts Representation
Alaa Alahmadi, Arash Joorabchi and Abdulhussain E. Mahdi
Electronic and Computer Engineering Department, University of Limerick, Limerick, Ireland
Keywords: Automatic Text Classification, Arabic Text, Bag-of-Words, Bag-of-Concepts, Wikipedia.
Abstract: With the exponential growth of Arabic text in digital form, the need for efficient organization, navigation
and browsing of large amounts of documents in Arabic has increased. Text Classification (TC) is one of the
important subfields of data mining. The Bag-of-Words (BOW) representation model, which is the
traditional way to represent text for TC, only takes into account the frequency of term occurrence within a
document. Therefore, it ignores important semantic relationships between terms and treats synonymous
words independently. In order to address this problem, this paper describes the application of a Bag-of-
Concepts (BOC) text representation model for Arabic text. The proposed model is based on utilizing the
Arabic Wikipedia as a knowledge base for concept detection. The BOC model is used to generate a Vector
Space Model, which in turn is fed into a classifier to categorize a collection of Arabic text documents. Two
different machine-learning based classifiers have been deployed to evaluate the effectiveness of the
proposed model in comparison to the traditional BOW model. The results of our experiment show that the
proposed BOC model achieves an improved performance with respect to BOW in terms of classification
accuracy.
1 INTRODUCTION
Arabic, which is the fifth most spoken language in
the world, belongs to the Semitic group of languages
and it is a highly inflectional and derivational
language (Versteegh and Versteegh, 1997). The
Arabic alphabet consists of 28 letters, and letters are
directly connected when forming words. Arabic text
is read from right to left. Unlike English, proper
nouns do not start with capital letters which makes
the process of recognizing and extracting them
difficult. Moreover, 22 of the letters take different
shapes based on their position in the word, i.e.,
initial, medial, or final. Table 1 shows the letter b
() appearing in different sample words and how its
position affects its shape.
Table 1: Example of an Arabic letter and its various
shapes depending on its position in the word.
position word Letter shape
Start
Middle
End
The exponential growth of Arabic documents in
digital form on the web has increased the need to
assist users with the fast and effective navigation,
browsing, and discovery of useful information on
the Internet. Text Classification (TC) is the task of
assigning one or more predefined categories to a
given text.
Classification of Arabic text can be a challenging
task due to the rich and complex nature of the
Arabic language. The majority of reported works on
Arabic TC attempt to represent text by using the
Bag-of-Words (BOW) model. In this model, each
document,d
∈D, is expressed as a weighted high
dimensional vector, d
i
, where each dimension
corresponds to a unique word. For example,
Alsaleem (2011) used Support Vector Machine
(SVM) and Naïve Bayes (NB) classification
algorithms using the BOW model to classify a Saudi
Arabic newspaper text collection by Al-Harbi et al.
(2008). The classification system yielded a Macro
F1 score of 77.85% and 74.0% for SVM and NB
respectively. Kanaan et al. (2009) used the BOW
model for Arabic TC using a dataset compiled by
Mesleh (2007) with different weighting schemes
such as Term Frequency (TF), Term Frequency
Inverse Document Frequency (TFIDF), and
Weighted Inverse Document Frequency (WIDF).
374
Alahmadi A., Joorabchi A. and E. Mahdi A..
Arabic Text Classification using Bag-of-Concepts Representation.
DOI: 10.5220/0005138103740380
In Proceedings of the International Conference on Knowledge Discovery and Information Retrieval (KDIR-2014), pages 374-380
ISBN: 978-989-758-048-2
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
Performance comparisons between k-Nearest
Neighbors (k-NN), Rocchio, and NB as classifiers
were conducted. Results of this experiment show
that the WIDF scheme yields the best performance
when used in conjunction with the k-NN, while
TFIDF shows the best performance when used in
conjunction with Rocchio. Among the three
classifiers, the NB classifier is reported to be the
best performer yielding a Macro F1 score of
84.53%. Mesleh (2007) reported a BOW based
Arabic TC system which uses Chi-square for feature
selection. Their results show that using a SVM
classifier in this context yields better classification
performance compared to a k-NN or a NB classifier
when features are reduced using Chi-square. It
yields a Macro F1 score of 88.11%, when evaluated
using their in-house compiled Arabic text dataset.
The BOW model suffers from two main
limitations: (1) it breaks terms into their constituent
words, e.g., it breaks ‘text classification’ into the
words ‘text’ and ‘classification’; as a result the order
of the words is lost in the model and the meaning of
the terms could be changed; (2) it treats synonymous
words as independent features, e.g., ‘classification’
and ‘categorization’ are considered as two
independent words with no semantic association. As
a result of that documents which discuss similar
topics and contain synonymous words could be
considered unrelated.
Researchers have attempted to address the above
issues in English TC by representing text as
concepts rather than words, using an approach
known as Bag-of-Concepts (BOC). A concept is a
unit of knowledge with a unique meaning (ISO,
2009). To build a BOC model, semantic knowledge
bases such as WordNet
1
, Open Directory Project
(ODP)
2
, and Wikipedia
3
are used to identify the
concepts appearing within a document. By using
concepts in text representation the semantics and
associations between words appearing in the
document will be preserved. For example, Hotho et
al (2003) used the English Wordnet as a knowledge
base to represent English text. For each term in a
document, Wordnet returns an ordered list of
synonyms and the first ranked synonym will be used
as a concept for the term. In that study, three
approaches were proposed for using concepts as
features for text representation; (1) using only
concepts to represent documents; (2) Adding
Concept (AC) as complimentary features to the
1
http://wordnet.princeton.edu/.
2
http://dmoz.org.
3
http://www.wikipedia.org.
BOW model ; (3) Replacing Term words with
Concepts (RTC) in BOW model. The study shows
that AC yields better classification performance
results than the other two approaches. The results of
this study show that representing documents with
only concepts is insufficient as WordNet does not
cover all special domain vocabularies. Furthermore,
WordNet is limited as it is a manually constructed
dictionary and therefore laborious to maintain.
To deal with this problem, other researchers have
tried replacing WordNet with other knowledge bases
derived from the Internet, such as ODP and
Wikipedia, e.g., see (Gabrilovich and Markovitch,
2005, Gabrilovich and Markovitch, 2006). In these
studies, ODP categories and Wikipedia articles are
used as concepts for text representation. For each
document, a text fragment (such as word, sentence,
paragraph, or the whole document) maps to the most
relevant ODP categories or Wikipedia articles. The
mapped concepts are added to the document using
the AC approach. Using these knowledge bases for
BOC modelling improved the performance when
applied to English TC compared to BOW model.
For Arabic TC, Elberrichi and Abidi (2012) used
the Arabic WordNet (Black et al., 2006) to identify
concepts appearing within the documents. A
comparison between different text representation
models such as BOW, N-grams and BOC was
conducted using an Arabic text dataset collected by
Mesleh (2007). An RTC variation of BOC used in
conjunction with Chi-square for feature selection
and a k-NN classifier was reported to achieve higher
performance results compared to other
representations.
In a previous work, we developed a number of
new approaches, which combine the BOW and the
BOC models, and applied them to English TC
(Alahmadi et al., 2013) and Arabic TC (Alahmadi et
al., 2014). In (Alahmadi et al., 2014), a NB
classification algorithm was shown to provide
better performance in conjunction with the BOC
model compared to BOW model. This current work
focuses on this point. The remainder of the paper is
organized as follows: Section 2 describes why
Wikipedia is a suitable knowledge base for BOC
modelling of Arabic text. Section 3 outlines the pre-
processing phase and describes the BOC model in
details. The experimental set-up and results are
discussed in Section 4. The paper concludes in
Section 5.
2 WIKIPEDIA
Wikipedia is the largest electronic knowledge
ArabicTextClassificationusingBag-of-ConceptsRepresentation
375
repository on the Internet and its content is entirely
contributed collaboratively by volunteers. Wikipedia
is a comprehensive, up to date, and well-formed
knowledge source.
In order to use Wikipedia as a knowledge source
for building BOC models of Arabic text, an open-
source toolkit known as Wikipedia-Miner (Milne
and Witten, 2013) is utilized. First the toolkit
processes Wikipedia XML dump files
4
and creates a
database that contains a summarized version of
Wikipedia’s content and structure. The toolkit
assigns each Wikipedia page a unique id, title, and
type (article, category, or redirect). Each article in
Wikipedia describes a single concept and the article
title is the descriptor of the concept. The title is well
formed, brief and can be used as a descriptor in
ontologies. The aim of the redirect pages in
Wikipedia is to connect articles with alternative
titles that correspond to their synonyms.
The great majority of Wikipedia articles, i.e.
concepts, are classified to one or more Wikipedia
categories, which are hierarchical and descend from
a single root. The maximum category depth in
Arabic Wikipedia is 12. A concept can belong to
multiple parent categories which are more general
articles than the concept itself. In addition more
specific concepts which are child articles can be
mined by the concept as a parent category for them.
In Wikipedia, an article may have more than one
redirect and other Wikipedia articles may link to it
(link anchors). All these elements offer additional
information, and they are grouped into labels. These
labels will be used as concepts to represent text in
different applications. In this study we have used
Wikipedia labels to build the BOC representation
model for Arabic TC.
3 ARABIC TEXT
CLASSIFICATION WITH BOC
MODEL
This section describes the developed TC system as
shown in Figure 1.
The process begins by passing all the documents in
the dataset through a text pre-processing phase
where they are cleaned and processed. The result of
this phase is a set of well-defined features. These
features are then used in the modeling phase. Based
on the features, i.e. words or concepts, two
independent representation models are built; namely
4
http://dumps.wikimedia.org/arwiki/
the BOW and BOC. Finally, the outputs of these
phases are fed to two distinct Machine Learning
(ML) based classification algorithms (classifiers).
These algorithms learn from the labeled training
texts to predict the class of unlabeled testing texts.
The remaining of this section elaborates on each of
these phases.
Dataset
Pre‐
processing
Modelling
BOW
Classification
algorithms
BOC
Figure 1: Developed TC system.
3.1 Text Pre-Processing
The pre-processing is the task of converting text to a
well-defined set of features. In this phase all noise
and irregularities, which negatively affect the
classification performance, will be removed.
This phase includes the following steps:
a. All digits, punctuation marks, and non-Arabic
characters are removed.
b. For normalization we follow (Kanaan et al.,
2009, Mesleh, 2007) and remove diacritics and
normalize some Arabic letters as follows:
Normalization of () by replacing (), () and
() at the start of words with ().
Normalization of () by replacing () at the
end of words with ().
Normalization of () by replacing () at the
end of words with ().
c. Arabic stop words and common words such as
pronouns and prepositions are removed from
text as they do not carry any discriminatory
significance so far classification is concerned.
d. There are two slightly different approaches to
stemming in Arabic language: (a) Root
Extraction (RE) in which a set of prefixes,
suffixes, and infixes are removed from words to
extract root; and (b) Light Stemming (LS) in
which only prefixes and suffixes are removed
from words. Prefixes in Arabic words are
groups of letters added to the beginning of the
root, these letters could be definite articles,
prepositions, pronouns and connectives ,such as
(), (), (), () and (). Suffixes in
Arabic words are set of letters that are added to
the end of the root, such as (), (), (), ( ),
( ), () and (). Infixes in Arabic words are
set of letters that inserted into root to create new
words; such as (), (), (
), () and ().
In this work, LS approach has been used as it
has been found to yield better results compared
KDIR2014-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval
376
to RE for Arabic TC (Kanaan et al., 2009,
Mesleh, 2007).
e. Remove words that occur less than 4 times in
the dataset to reduce the dimensionality of the
feature vector of the representation model
(Mesleh, 2007).
3.2 Text Representation
The result of pre-processing phase is a set of features
that represent Arabic text. In our classification
system features can be either words or concepts and,
therefore two independent representation models are
created, i.e., BOW and BOC. By using words as
features, a BOW model is created and each word is
weighted using the TFIDF weighting scheme (Salton
and Buckley, 1988):
TFIDF
w
,d
TF
w
,d
∙IDFw
(1)
where TF
w
,d
is the frequency of a word,w
∈
W, W is the set of all words that are considered as
features.
TF
w
,d
nw
,d
|
d
|
(2)
where nw
,d
is the occurrence frequency of the
word w
in documentd
, normalized by
|
d
|
which
is the length ofd
. The inverse document frequency
IDF
w
is defined as:
IDF
w
log
|
D
|
DFw
(3)
where DF
w
returns document frequency of the
word w
. It counts the number of documents in D
where the word w
appears,
|
D
|
is the total number
of documents in the dataset. The IDF
w
parameter has the effect of reducing the weight of
those words which appear in a large portion of the
datasetD.
To create the BOC model, concepts within a
given text need to be identified using the Wikipedia-
Miner toolkit. First, a document is processed and
cleaned. Then, all possible Wikipedia concepts in
the document will be identified using the topic
detection functionality in Wikipedia-Miner. A
majority of detected concepts are ambiguous as they
may refer to multiple meanings. The disambiguation
component of the toolkit is used to compute a
probability estimate for each potential concept based
on its relevance to other detected concepts in the
document. The probability determines if a concept is
relevant to the context or not and the ambiguity will
be resolved. A set of candidate concepts are selected
which are used to represent documents in the BOC
model.
For two given documents d
1
and d
2
which belong to
SNP dataset (see Section 4 for the details of this
dataset); the first document belongs to ‘Sport’
category and describes a news article about Real
Madrid football team, whereas the other document
belongs to the ‘Economic’ category and discusses
the Saudi stock market. The word Riyal () is
appeared in both documents but with different
intended semantic meaning. In d
1
it refers to Real
Madrid ( ) team and in d
2
it refers to Saudi
Riyal ( ) currency. By using BOW
representation the difference in the meanings are lost
and the Riyal () is considered the same in both
documents. Furthermore, as can be seen in Table 2,
the value for Riyal () document frequency in the
dataset is 267 which reflects the generality of this
feature. This leads to reducing the TFIDF weight of
this feature and as a result the document is classified
to a wrong category.
Table 2: Sample of features from SNP dataset and their
Document Frequency (DF) values.
Features
DF
Riyal
()
Madrid
()
Saudi
()
Saudi
Riyal
(
)
Real
Madrid
(
)
267 23 581 179 9
However, by using the BOC model various
meanings of the terms will be identified and mapped
to their corresponding concepts, e.g., Riyal ()
such as Saudi Riyal ( ), Omani Riyal )
(, Qatari Riyal ( ), Yemeni Riyal )
( and Real Madrid ( ), and based on the
context of the document most related concepts are
selected.
Table 3: Sample of weighted features in documents d
1
and
d
2
with BOW representation model.
Features
d
1
d
2
TF TFIDF TF TFIDF
Riyal () 2 0.2568 1 0.01098
Saudi () 2 0.1892 0 0
Madrid () 0 0 1 0.0200
Table 3 shows a sample of features with their
corresponding weights in both sample documents d
1
and d
2
with the BOW model. This example
demonstrates how the word Riyal () and Madrid
() have low weights in document d
2
despite their
importance in the classification step. Table 4 shows
that Real Madrid ( ) weight is increased in
ArabicTextClassificationusingBag-of-ConceptsRepresentation
377
BOC model and reflect the importance of the
concept in document d
2
.
Table 4: Sample of weighted features in two documents in
BOC representation model.
Features
d
1
d
2
TF TFIDF TF TFIDF
Saudi Riyal
( )
2 0.4374 0 0
Real Madrid
( )
0 0 1 0.1359
3.3 Classification Algorithms
In the last twenty years a wide range of ML
algorithms have been used for TC. The ML
algorithms automatically learn from a set of pre-
classified (labelled) documents.
In this work two classification algorithms have been
used, namely NB and Random Forest.
a. Naive Bayes Classifier: Naive Bayes (NB) is a
probabilistic classifier based on applying Baye's
theorem, and is commonly used in ML applications
(Mitchell, 1996). The basic idea in a NB-based
classifier is to estimate the probabilities of categories
for a given document by observing the joint
probabilities of features and categories.
b. Random Forest Classifier: Random Forest is
a commonly used classification method and it
proposed by Breiman (2001). Random Forest builds
a set of classification trees based on a subspace of
features randomly selected to predict a category of a
text instance.
4 EVALUATION
To evaluate the performance of the BOC
representation
model, we have assessed the accuracy
of the developed classification system using two
Arabic text classification datasets:
a. Arabic 1445 Dataset: This dataset has been
collected by Mesleh (2007) from online Arabic
newspaper archives including Al-Jazeera, Al-Nahar,
Al-hayat, Al-Ahram, and Al-Dostor. The dataset
contains 1445 documents that vary in length and fall
into nine categories: Computer, Economics,
Education, Engineering, Law, Medicine, Politics,
Religion, and Sports.
b. Saudi Newspapers (SNP) Dataset: The
dataset consists of 5121 Arabic documents of
different lengths which belong to seven categories:
Culture, Economics, General, Information
Technology, Politics, Social, and Sport. It has been
collected by Al-Harbi et al. (2008) and consists of
articles and news stories from Saudi newspapers.
We conducted all the experiments using WEKA
(Hall et al., 2009), which is a popular open source
toolkit for machine learning. We first converted the
textual documents into the format required by
WEKA, i.e., ARFF format (Attribute-Relation File
Format)
5
. We then used this data to train two
separate classification algorithms namely, NB and
Random Forest. This was then followed by a ten-
fold cross validation to test and evaluate the
performance of the classifiers, in terms of Accuracy
and Macro F1 measures. These performance
measures are built upon the concepts of Precision
(Pr) and Recall (Re). Precision is the probability that
a document predicted to be in categoryc
, truly
belongs to this category. Recall is the probability
that a document belonging to c
is classified into this
category. When a single performance measure is
desired, the harmonic mean of the precision and
recall, F1, is quoted. The Accuracy (Acc) is
computed by dividing the total number of documents
assigned to a given categoryc
by the total number
of documents in the testing dataset. Let C = (c
1
, .., c
i
)
denote the set of categories in the dataset,
accordingly with respect to a given categoryc
:
Re
c
TP
TP
FN
(4)
Pr
c
TP
TP
FP
(5)
F1
c
2Prc
Rec
Pr
c
Rec
(6)
Acc
c
TP
TN
TP
TN
FP
FN
(7)
where, the Re, Pr , F1 and Acc are computed in
terms of the labels TP (True Positive), TN (True
Negative), FP (False Positive) and FN (False
Negative), such that:
TP
i
: refers to the cases when both the classifier
and human cataloguer agree on assigning
category c
to document d;
TN
i
: refers to the cases when both the
classifier and human cataloguer agree on not
assigning category c
to document d;
5
http://www.cs.waikato.ac.nz/ml/weka/arff.html
KDIR2014-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval
378
FP
i
: refer to the cases when the classifier has
mistakenly (as judged by a human cataloguer)
has assigned category c
to document d;
FN
i
: refers to the cases when the classifier has
failed (as judged by a human cataloguer) to
assign a correct category c
to document d.
For automatic text classification, precision and
recall values for the various classes should be
combined to obtain an accurate measure of the
classifier algorithms used. This is commonly done
using the Macro-averaged performance, which is
calculated by first computing the scores per
category, i.e., Re (c
i
), Pr (c
i
), F1 (c
i
), and then
averaging these per-category scores to compute the
global means.
Figure 2 compares the NB-based classifier
accuracy achieved by the BOW model and BOC
model as applied to both Arabic datasets. Figure 3
on the other hand compares the same for the
Random Forest based classifier.
Figure 2: Ten-fold average accuracy achieved by the NB-
based classifier for the BOW and BOC text representation
models.
Figure 3: Ten-fold average accuracy achieved by the
Random Forest -based classifier for the BOW and BOC
text representation models.
Both Figures 2 and 3 indicate that the BOC
representation used in our experiments for Arabic
TC outperforms the BOW representation. The NB
classifier seems to yield a 5% higher accuracy
compared to the Random Forest classifier when
applied to the Arabic 1445 dataset, and a 3% higher
accuracy in the case of the SNP dataset. It has also
been noted that the BOC based classifiers offer
faster the execution time comapred to that of
BOW’s. This is due to the fact that the BOC model
tends to represent a document using fewer features
compared to the BOW model. Table 5 gives detials
of the number of features used in each text
represetnation model used in the NB classifer in our
experiements and corresponding execution times of
the classification algorithm in each case. In this
experiment a desktop PC with a 3.33GHz Pentium
processor operating under Windows 7 was used.
Table 5: Number of features and execution times for the
NB-based classifier for both text representation models.
Model Number
of features
Classifier
execution time (s)
BOW 12821 5.61
BOC 3463 0.96
Figure 4 shows the performance of the NB-based
classifier, in terms of the Macro F1 measure, for
both BOW and BOC models. The figure clearly
shows that the BOC model outperforms the BOW by
10% when applied to the SNP dataset. In contrast to
other classifiers that use the BOW model, our NB-
based classifier with the BOC model achieved
88.99% in Macro F1 measure when applied to the
Arabic 1445 dataset, as compared to 84.53%
achieved by NB-based classifier reported by Kanaan
et al. (2009), and 88.11% achieved by the SVM-
Figure 4: Performance of the NB-based classifier in terms
of the Macro F1 (ten-fold averaged) for both the BOW and
BOC text representation models.
ArabicTextClassificationusingBag-of-ConceptsRepresentation
379
based classifier reported by Mesleh (2007). For the
SNP dataset, our NB-based classifier utilising the
BOC model achieved 80.24% in Macro F1 measure
compared to 74.0% achieved by the same type of
classifier reported by Alsaleem (2011).
5 CONCLUSIONS
In this work, we demonstrated that Arabic text
classification can be improved by representing
textual documents as a set of concepts using the
BOC model. By doing so, background knowledge
can be introduced to the document representation
from sources such as Wikipedia to overcome some
of the limitations of the classic BOW representation.
As demonstrated in our reported experimental
results, the described BOC text representation model
significantly improves the classification accuracy
compared to the BOW model when evaluated using
two Arabic TC datasets.
REFERENCES
Al-Harbi, S., Almuhareb, A., Al-Thubaity, A., Khorsheed,
M. and Al-Rajeh, A. 2008. Automatic Arabic text
classification.
Alahmadi, A., Joorabchi, A. and Mahdi, A. E. A new text
representation scheme combining Bag-of-Words and
Bag-of-Concepts approaches for automatic text
classification. GCC Conference and Exhibition
(GCC), 2013 7th IEEE, 2013. IEEE, 108-113.
Alahmadi, A., Joorabchi, A. and Mahdi, A. E. 2014.
Combining Bag-of-Words and Bag-of-Concepts
Representations for Arabic Text Classification. IET
Irish Signals & Systems Conference 2014.
Alsaleem, S. 2011. Automated Arabic Text Categorization
Using SVM and NB. Int. Arab J. e-Technol., 2, 124-
128.
Black, W., Elkateb, S. and Vossen, P. Introducing the
Arabic wordnet project. In Proceedings of the third
International WordNet Conference (GWC-06, 2006.
Citeseer.
Breiman, L. 2001. Random forests. Machine learning, 45,
5-32.
Elberrichi, Z. and Abidi, K. 2012. Arabic Text
Categorization: a Comparative Study of Different
Representation Modes. International Arab Journal of
Information Technology (IAJIT), 9.
Gabrilovich, E. and Markovitch, S. Feature generation for
text categorization using world knowledge. IJCAI,
2005. 1048-1053.
Gabrilovich, E. and Markovitch, S. Overcoming the
brittleness bottleneck using Wikipedia: Enhancing text
categorization with encyclopedic knowledge. AAAI,
2006. 1301-1306.
Hall, M., Frank, E., Holmes, G., Pfahringer, B.,
Reutemann, P. and Witten, I. H. 2009. The Weka data
mining software: an update. ACM SIGKDD
explorations newsletter, 11, 10-18.
Hotho, A., Staab, S. and Stumme, G. 2003. Wordnet
improves Text Document Clustering.
ISO. (2009). ISO-704: Terminology work—Principles and
methods (3rd ed.).
Geneva, Switzerland: International Organization for
Standardization.
Kanaan, G., Alshalabi, R., Ghwanmeh, S. and Alma'adeed,
H. 2009. A comparison of text classification
techniques applied to Arabic text. Journal of the
American society for information science and
technology, 60, 1836-1844.
Mesleh, A. M. D. A. 2007. Chi Square Feature Extraction
Based Svms Arabic Language Text Categorization
System. Journal of Computer Science, 3.
Milne, D. and Witten, I. H. 2013. An open-source toolkit
for mining Wikipedia. Artificial Intelligence, 194,
222-239.
Mitchell, T. 1996. Machine Learning. McCraw Hill.
Salton, G. and Buckley, C. 1988. Term-weighting
approaches in automatic text retrieval. Information
processing & management, 24, 513-523.
Versteegh, K. and Versteegh, C. 1997. The Arabic
Language, Columbia University Press.
KDIR2014-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval
380
