Classification of Arabic Information Extraction methods
Abd El Salam AL HAJJAR
Institute University of Technology
Lebanese University
Lebanon
Paragraph Laboratory
University of Paris 8- Vincennes- Saint-Denis
France
abdsalamhajjar@hotmail.com
Mohammad HAJJAR
Institute University of Technology
Lebanese University
Lebanon
m_hajjar@ul.edu.lb
Khaldoun ZREIK
Paragraph Laboratory
University of Paris 8- Vincennes- Saint-Denis
France
zreik@univ-paris8.fr
Abstract
The performance of information retrieval in arabic language is very problematic due to the specific morphological and structural
changes in the language. To extract information from an arabic document, the involved methods must answer the following question:
"How can we find the root of the word we search". To find a word in an arabic dictionary, first you must extract the root of this word
and then find this root in the dictionary. This is because the vocabulary of the arabic language is essentially built from the roots
derivation. The roots are words composed of three to five consonants letters. To address these problems, several methods have been
proposed. The aim of this paper is to propose a preliminary classification of arabic information extraction methods. These methods can
be classified into two main categories. The first one is called "Stemmer". This category includes the following subcategories: Stemmer
based on affixes, Stemmer based on translation and Stemmer based on pattern and affixes. The second is called "N-gram". This
category regroups the subcategories: N-gram based on Dice's similarity coefficient and N-gram based on “Manhattan distance”
dissimilarity coefficient. However, we find methods which implement the two approaches "Stemmer" and "N-gram". This work
contributes to decide on the more appropriate arabic information extraction method.
2005, Khoja & Garside, 1999; Hammo et al., 2002). The
second is called "N-gram". It based on statistical
approaches to retrieve the information independently of
the language complexity (Adamson George & Boreham,
1974; Suleiman Mustafa, 2004; Ahmed & Nürnberger,
2007; M. Sinane et al., 2008; Khreisat, 2006). However,
we find methods which implement the two approaches
"Stemmer" and "N-gram" (De Roeck & Al-Fares, 2000).
Introduction
Arabic language is used by more than 330 million arabic
speakers that are spread over 22 countries (Ghosn, 2003;
Censure of the Internet in the arab countries, 2006).
However, the performance of information retrieval in
arabic language is very problematic due to the specific
morphological and structural changes in the language:
polysemy, irregular and inflected derived forms, various
spelling of certain words, various writing of certain
combination character, short (diacritics) and long vowels,
most of the arabic words contain affixes (Table 1, 2). To
address these problems, several methods have been
proposed.
The aim of this paper is to propose a preliminary
classification of arabic information extraction methods.
These methods can be classified into two main categories.
The first one is called "Stemmer" which requires specific
knowledge about the language (Al Ameed et al., 2005;
Larkey et al., 2002; Larkey, 2005; Darwish, 2002; Chen &
Gey, 2002; Kanaan et al., 2004; Thabet, 2004; Kadri &
Nie, 2006; Al-Shalabi & Evens, 1998; Taghva et al.,
Problematic
To find a word in an arabic dictionary, first we must
extract the root of this word and then find this root in the
dictionary (Ibn Manzour, 2008). This is because the
vocabulary of the arabic language is essentially built from
the roots derivation. The roots are words composed of
three to five consonants letters. The arabic language has
about ten thousand roots, 85% of them are trilateral. The
derivation of words is done by adding affixes (prefix,
infix, or suffix) to the root according to several patterns
that are around 120 (Al Kharashi, 1999). For example, let
us take the root ( ); the words (
,
,
) (Table
71
3) are respectively derived from this root according to the
patterns (
,
,
) (Table 4).
To extract information from an arabic document, the
involved methods must answer the following question:
"How can we find the root of the word we search". To
answer this question, we must perform a morphological
analysis. In the arabic language, this consists to identify
the morphemes of a word (Stem): the affixes (prefix,
infix, and suffix) and the root. A stem can be a noun, verb
or particle. It can be composed of: One part (a root, for
example: (
)); Two parts (a root + a pattern, for
example: (
): root (
) + a pattern (CuCiC where
C is the consonants of the root (the radicals)); Three parts
(a root + a pattern + affixes, for example: (
ا
)ا:
root (
) + a pattern CaCiC + affixes (prefix (al) ( )ا,
infix (a) ( )اand the suffix ( a) ( ))) (Table 2, 3).
system. This category includes the following
subcategories: Stemmer based on affixes, Stemmer based
on translation and Stemmer based on pattern and affixes.
Stemmer based on affixes
Several Stemmer algorithms use the predefined rules to
remove the affixes (prefix, infix, suffix ...) from the word
to extract the root. This category allows remarkably good
information retrieval without providing correct
morphological analysis.
Several algorithms have been developed (Al Ameed et al.,
2005; Larkey et al., 2002; Larkey, 2005; Chen & Gey,
2002; Kanaan et al., 2004; Thabet, 2004; Kadri & Nie,
2006).
The normalization phase is applied before applying these
algorithms (Replacing “ ”أ, “ ” إand “ ”آby alif bar “”ا,
Replacing “ ” by “ ”يat the end of the words, Replacing
“ ” by “ ” at the end of the words, Replacing the sequence
“ ” يءby “”ئ, etc.) (Table 1).
Al Ameed, H. et al. (2005) and Larkey, L. et al.(2005)
developed a light stemmer which is based on the
suppression of “ ” if it is initial at the beginning of the
word, of the prefixes (
, , , , ا, )ا, and of the
suffixes ( , ي, , ا, ا, ي, , , يي, )ي. A. Chen and F.
Gey (2002) identified other sets of prefixes and suffixes.
To remove the prefixes and suffixes in the pre-defined
sets, each algorithm proposes their own rules. For
example, A. Chen and F. Gey (2002) apply the following
rules: If the word is at least five-character long, remove
the prefixes of three characters: , اا, , , , , , ا
ا, . If the word is at least four-character long, remove
the first two characters: ا, ا, ,
, ي, ا, , , , ,
, , , ي. If the word is at least four-character long and
begins with remove the initial letter . If the word is at
least four-character long and begins with either
or
remove or (Table 2).
Kanaan et al. (2004) presented a new stemming algorithm
to extract quadrilateral arabic roots. The algorithm starts
by excluding the prefixes, and then checks the word
characters starting from the last letter backward to the first
one. A temporary matrix is used to store the suffix letters
of the arabic word, and another matrix is used to store the
roots. Algorithm checks the letters of any word, also
checking whether the tested letter is included within the
general standard arabic word.
The large-scale use of diacritics ( , , , , , , , ) (Table 1)
representing short vowels are prevalent in the Qur’an.
Every word, even every letter is marked with a diacritic.
(For example:
"reign",
“king" ...). N. Thabet
(2004) proposes a new stemming approach based on a
light stemming technique that uses a transliterated version
of the Qur’an in western script (Table 3).
Y. Kadri and J. Nie (2006) defined that the arabic words
are usually formed as a sequence of antefixes are
generally prepositions joined to words at the beginning
( ..., , , , ا, ), prefix are usually represented by
only one letter and indicate the conjugation person of
verbs in the present tense ( ... , ي, ,( )اTable 1), core,
suffixes are the conjugation terminations of verbs and they
are the dual/plural/female marks for the nouns ( , ي,
... ا, ا, , ) ي, and postfixes represent pronouns
attached to the end of the words(... , , ) ) (Table 2).
Results
The study that we realized permits to identify several
methods which address the problems of information
extraction from arabic documents. We have found that
these methods can be classified into three categories:
"Stemmer", "N-gram", and "Stemmer and N-gram". The
first category requires specific knowledge about the
language. The second is based on statistical approaches to
retrieve the information independently of the language
complexity. In the last category we find methods which
implement the two approaches "Stemmer" and "N-gram".
The diacritics and the variation of the letter forms
according to its positions take an important role in the
arabic reading and writing complexity and reduce the
Arabic Information Extraction methods performance. To
resolve these problems, the normalization phase is applied
before applying these methods, the text normalization
takes a character string as input and tries to remove or
replace some characters under the predefined rules to
converts it into a string of letters (Figure 1). Every
method has the specific rules, in general a text is
normalized by removing (the tatweel character “-“, the
diacritics and the shedda “ ”, the punctuations, the non
letters, the stop words, the specials characters, and the
numbers) and replacing (“”أ, “ ”إand “ ”آby alif bar “”ا, “
” by “ ”يat the end of the words, “ ” by “ ” at the end
of the words, “ ” يءby “”ئ, ؤby ء, and ….) (Chen, A.
& Gey, 2002; Kadri & Nie, 2006; Khreisat, 2006; Larkey
et al., 2002; Larkey, 2005; Douzidia & Lapalme, 2005;
Darwish, 2002).
Figure 1: Normalization Process
Stemmer Category
Stemmer is an automatic process used to reduce the
different morphological forms of words into common root
(Stem) to improve the performance of the extraction
72
N-gram Category
Stemmer based on translation
The algorithms of english Stemmer have better
performance than arabic. For that, several methods use the
translation technique to allow any languages (like arabic)
use the Stemmer of another language (like english) to
extract the root of a word. A. Chen and F. Gey (2002)
have built a MT-based arabic stemmer from the arabic
words found in the arabic documents, and their english
translations are using the online “Ajeeb” machine
translation system. They divided the arabic words into
clusters based on the english translations of the arabic
words. The arabic words whose english translations, after
removing english stop words, are conflated to the same
english stem that made from one cluster. All the arabic
words in the same cluster are conflated to the same arabic
word, which is the shortest arabic word in the cluster. For
example:
( أour children), remove "our" is a word
parasite,
أis apparent that in relation to "child". So
أis related to
(Table 3).
The second category is based on statistical approaches: Ngram. The N-gram can find if two words are semantically
similar or dissimilar from the structures of characters of
these words. Two words are considered similar if they
have in common several substring of N characters, this is
done by calculating a coefficient on these two words. The
advantages of N-gram are that it does not require a
preliminary knowledge of the language, does not require
predefined rules, and does not require the construction of
a database of vocabulary. This class gives a good result in
multiple languages, even in arabic (3-gram and 4-gram).
This category regroups the subcategories: N-gram based
on Dice's similarity coefficient and N-gram based on
“Manhattan distance” dissimilarity coefficient. In this
category you can find the following classes:
N-gram based on the calculation of similarity
coefficient
W. Adamson George and J. Boreham (1974) has been
developed the first automatic classification technique
based on the character structure of words. Dice's
Similarity Coefficient is computed from the number of
matching bigrams (2-gram) in pairs of character strings,
and used to cluster sets of character strings (Figure 2).
H. Suleiman Mustafa (2004) assesses the performance of
two N-gram matching techniques for arabic root-driven
string searching: contiguous N-grams and hybrid Ngrams, combining contiguous and non-contiguous.
F. Ahmed and A. Nürnberger (2007) have been presented
the n-gram model which can be used to compute the
similarity between two strings by counting the number of
similar n-grams they share. The more similar n-grams
they found between the two strings exist the more similar
they are. Based on this idea the similarity coefficient can
be derived. The similarity coefficient δ is defined by the
following equation:
Stemmer based on pattern and affixes
Several Stemmer algorithms based on the patterns and
affixes have been developed, to find roots with three
letters, four and five letters, starting from verbal forms,
nouns and adjectives derived from verbs (Khoja &
Garside, 1999; Hammo et al., 2002; Al-Shalabi & Evens,
1998; Taghva et al., 2005; Darwish, 2002).
S. Khoja and R. Garside (1999) have proposed a method
that
involves
removing
diacritics
representing
vowelization, the stop words, the punctuation, the
numbers, the definite article ( )ا, the inseparable
conjunction ( ), and the longest prefix and suffix. Then,
the result is compared to a list of patterns. If a match is
found, the characters representing the root in the pattern
are extracted (Table 1, 2).
The QARAB system is developed by B. Hammo et al.
(2002) is based on the Khoja Stemmer (Khoja & Garside,
1999).
R. Al-Shalabi and Mr. Evens (1998) proposed an
approach of Stemmer where the first step is to remove the
longest possible prefix. The three letters of the root must
lie somewhere in the first four or five characters of the
remainder. They checked all the possible trigrams within
the first five letters of the remainder. Then, they check the
following six possible trigrams: First, second, and third
letters. First, second, and fourth. First, second, and fifth,
etc.
K. Taghva et al. (2005) implemented a root-extraction
stemmer for arabic which has a performance equivalent to
the Khoja stemmer (Khoja & Garside, 1999). To
implement this algorithm, they have defined several sets
of the affixes (D diacritic: ... , , , , P3 prefix of
length 3: ا, . P2 prefix of length 2: , ا. P1 prefix of
length 1:
, , . S3 suffix of length 3:
,
, . S2
suffix of length 2: ا, ا, . S1 suffix of length 1: , ,
()يTable 1, 2) and several sets of models (PR4 model of
length 4: ,
,
. PR53 model of length five and a
root of length 3: ... أ, ا,
. PR54 model of length
five and a root of length four:
, ا,
. PR63
model of length 6 and a root of length 3:
ا
...
ا. PR64 model of length 6 and a root of length 4:
...
, ( )ا اTable 4). The extraction of the root in this
algorithm is based on the length of the normalized stem.
Where α and β are the n-gram sets.
For Example: shows an example of two arabic words:
ا ي
اإand ا
( اTable 3):
ا ا
ي
ي
ا ا
إ
إ إ
ا
Figure 2: Bigram similarity measure between two words
ا ي
اإand ا
ا
M. Sinane et al. (2008) presented an approach that uses Ngram based on the word and characters. Four basic types
have been explored, sometimes separately and sometimes
in combination: Word, lexical root, root, and N-gram. In
general, N-grams based on the stems are better than those
based on words, because the N-grams based on words
could have prefixes and suffixes which make more
mistakes in the Similarity between the document and
query.
73
N-Gram based on the Frequency Statistics technique
L. Khreisat (2006) presented the N-Gram Frequency
Statistics technique for classifying arabic text documents.
The technique employs a dissimilarity measure called the
“Manhattan Distance”, and “Dice’s measure”. A corpus
of arabic text documents was collected from online arabic
newspapers, 40% of the corpus was used as training
classes and the remaining 60% of the corpus was used for
classification. All documents, whether training documents
or documents to be classified went through a
preprocessing normalization phase that remove the
punctuation marks, the stop words, the diacritics, and the
non letters. For the training documents, the N-gram (N=3)
(the trigrams of the word ي
اare: ي, ي,
, , )ا
(Table 2) was generated for each document and saved in
text files. Then for each document to be classified, the Ngram frequency profile was generated and compared
against the N-gram frequency profiles of all the training
classes. This comparison is done by calculating Manhattan
distance and Dice’s measure.
A. N. De Roeck and W. Al-Fares (2000) presented a
method for arabic words sharing the same root. To
implement this method "Clustering Algorithm", depends
on two stages which is called "Two-Stage". In first step,
they applied the Light Stemming to remove affixes, the
second step is based on the Adamson algorithm with some
modifications. Each bi-gram assigned a weight (0.25 for
bi-gram containing low letter, 0.5 for bi-gram containing
the non-low letter, 1 for all other bi-gram).
Conclusion
In this article we have proposed a first classification of
arabic information extraction methods into three
categories: Stemmer, N-gram, and "Stemmer" and "Ngram". In the stemmer category we find the following
subcategories: Stemmer based on affixes, Stemmer based
on translation, and Stemmer based on pattern and affixes.
In the N-gram category we find the following
subcategories: N-gram based on Dice's similarity
coefficient and N-gram based on “Manhattan distance”
dissimilarity coefficient. However, we find a method
which implements the two approaches "Stemmer" and "Ngram". The next step will be the making of a detailed
comparative study of the early described categories. This
study will cover mainly the following topics:
performances, stabilities, usability, advantages, and
disadvantages. Another possible extension of the present
work is to test these categories in similar conditions.
These studies and tests will permit to designate the more
appropriate arabic information extraction method or to
propose a new one.
Category: Stemmer and N-gram
Each of the two approaches has advantages and
disadvantages, as long as the Stemmer approach depends
on the language, and its morphological complexity, and
always does not give the best performance... and the
statistical approach N-gram is independent of language
but has drawbacks in terms of synonyms. To do this, there
are authors trying to merge the two approaches, in order to
have a good method.
Letter
~
ء
ا
إ
أ
آ
Transcription
Tanween Fatha
Tanween Dama
Tanween Kasra
Fatha
Dama
Kasra
Sokon
Shedda
Maada
Hamza
Alef
Alef with Hamza on bottom
Alef with Hamza on top
Alef with Maada
Baa
Taa Marbouta
Taa
Tha
Jeem
H'a
Khaa
Dal
Zain
At
Begin
ـ
X
ـ
ـ
ـ
ـ
ـ
Writing
In
Middle
ــ
X
ــ
ــ
ـ ـ
ـ ـ
ـ ـ
ـ
ـ
At
End
ـ
X
ـ
ـ
ـ
ـ
ـ
ـ
ـ
Letter
ؤ
ي
ئ
Transcription
Raa
Thal
Seen
Sheen
Saad
Daad
T'aa
Zha
Ain
Jain
Faa
Qaf
Kaf
Lam
Meem
Noon
Haa
Waw
Hamza on waw
Alif Makzora
Yaa
Hamza on yaa
At
Begin
ـ
ـ
ـ
ـ
ـ
ـ
ـ
ـ
ـ
ـ
ـ
ـ
ـ
ـ
ـ
X
X
يـ
ئـ
Writing
In
Middle
ـ
ـ
ـ ـ
ـ ـ
ـ ـ
ـ ـ
ـ ـ
ـ ـ
ــ
ــ
ــ
ــ
ــ
ــ
ــ
ــ
ــ
ـ
ـ
X
ـيـ
ــ
At
End
ـ
ـ
ـ
ـ
ـ
ـ
ـ
ـ
ـ
ـ
ـ
ـ
ـ
ـ
ـ
ـ
ـ
ـ
ـ
Table1: Arabic diacritics and letters transcription. Empty case means no writing change in the corresponding letter and
position. X-case means no existing of the corresponding letter
74
ـي
ـئ
Affix
اا
ا
ا
ا
ي
ي
ا
ا
إ
ا
ي
ي
Transcription
Alef Alef Lam
Alef Taa
Alef Lam
Alef Noon
Baa Alef Lam
Taa Meem
Dal Ain Yaa
Saa Sokon
Saa Alef Meem
Ain Yaa Noon
Faa Alef Lam
Kaf Alef Lam
Kaf Noon
Lam Alef
Lam Alef Lam
Lam Alef with Hamza on bottom
Lam Lam
Meem Alef Lam
Meem Waw Dal
Haa Alef
Haa Meem Alef
Haa Meem Lam
Waw Alef Lam
Waw Baa
Waw Taa
Waw Dal Ain
Waw Seen
Waw Lam Lam
Waw Meem
Waw Noon
Waw Yaa
Yaa Noon
Affix
ي
ا
ا
ا
ي
ي
ا
يي
ا
إ
ا
ي
ا
ي
ي
Transcription
Yaa Waw Noon
Alef Taa
Alef Lam Meem
Alef Noon
Baa Alef
Baa Alef Lam
Taa Alef Noon
Taa Meem Alef
Taa Yaa Noon
Seen with Fatha
Seen Yaa
Faa Alef
Kaf Alef Lam
Kaf Meem Alef
Waw Alef Lam
Waw Baa Alef Lam
Yaa Yaa Taa Marbouta
Alef Raa
Alef with Hamza on bottom Seen
Taa Meem
Taa Meem Lam
Raa Alef
Raa Yaa
Seen with Dama
Seem Taa
Kaf Alef Lam
Meem Raa
Waw Alef
Waw Lam
Waw Noon
Yaa Taa Marbouta
Yaa Haa
Table2: Arabic Affix Transcription cited in this paper and their transcription
Word
ا
ا
ا
ا
ا
ا
ا ي
ي
أ
اإ
ا
ا
Transcription
Kateb
Aleidwan
Alharb
Almaaraka
Kataba
Estemrar
Tofol
Kateba
Malek
Maktob
Estemrareya
Al modeoon
Atfalona
Muluk
Alkateboun
Translation
Writer
Attack
War
Battle
Write
Continuity
Child
Writer
King
Written
Continuities
Depositors
Our children
Had
Writers
Table3: Arabic words cited in this paper, their transcription, and their translation
75
Pattern
ا
ا
أ
Transcription
Eftaala
Afaalal
Afaal
Faaela
Faaol
Mafool
Efteaal
Moftaeel
Mostafeel
Mafaala
Mafaalal
Iestafaal
Afaalal
Tafaool
Tafaalal
Faeel
Faaela
ا
ا
ا ا
Table4: A sample of arabic pattern cited in this paper and their transcription
identifying Arabic roots". In Proceedings ACL-2000.
Hong Kong.
Douzidia, F. & Lapalme, G. (2005). "Un système de
résumé de textes en arabe", 2ème Congrès International
sur l'Ingénierie de l'Arabe et l'Ingénierie de la langue,
Alger.
Ghosn, Z. (2003). government Web sites in the Middle
East in 2003, The Arab Advisors Group.
Hammo, B. & Abu-Salem, H. & Lytinen, S. & Evens, M.
(2002). "A Question Answering System to Support the
Arabic Language". Proceedings of the ACL-02
workshop on Computational approaches to Semitic
languages Philadelphia, Pennsylvania Pages: 1 – 11.
Ibn Manzour (2008). Lisan Al-Arab. www.muhaddith.org.
Kadri, Y. & Nie, J. (2006). "Effective Stemming for
Arabic Information Retrieval" in proceedings of the
Challenge of Arabic for NLP/ MT Conference,
Londres, Royaume-Uni.
Kanaan, G. & Al-Shalabi, R. & Jaarn, J. & Al-Kabi, M. &
Hasnah, A. (2004). "A New Stemming Algorithm to
Extract Quadri-Literal Arabic Roots".
Khoja, S. & Garside, R. (1999). "Stemming Arabic text",
Computing
Department,
Lancaster
University,
Lancaster,
ww.comp.lancs.ac.uk/computing/users/khoja/stemmer.p
s, 1999.
Khreisat, L. (2006). "Arabic Text Classification Using Ngram Frequency Statistics a Comparative Study". The
2006 International conference on Data Mining Part of
the 2006 World Congress in Computer Sciences DMIN
2006: 78-82.
Larkey, L. S. & Ballesteros, L. & Connel, M. E. (2002).
"Improving Stemming for Arabic Information
Retrieval: Light Stemming and Co-occurrence
Analysis", in Proc. of the 25th annual international
ACM SIGIR conference on Research and development
in information retrieval, pp. 275 – 282.
Acknowledgements
This work has been done as a part of the project “Arabic
Web Intelligence” supported by the Lebanese National
Centre of Scientific Research (CNRSL).
Bibliographical References
Adamson George, W. & Boreham, J. (1974). "The use of
an association measure based on character structure to
identify semantically related pairs of words and
document titles", Information Storage and Retrieval,
Vol. 10, pp 253-260, 1974.
Ahmed, F. & Nürnberger, A. (2007). “N-grams
Conflation Approach for Arabic”, ACM SIGIR
Conference, Amsterdam, 27 July.
Al Ameed, H. & Al Ketbi, S. & Al Kaabi, A. & Al Shebli,
K. & Al Shamsi, N. & Al Nuaimi, N. & Al Muhairi, S.
(2005). "Arabic Light Stemmer: A new Enhanced
Approach", The Second International Conference on
Innovations in Information Technology (IIT’05).
Al Kharashi, I. (1999). "A Web Search Engine for
Indexing,
Searching
and
Publishing
Arabic
Bibliographic Databases".
Al-Shalabi, R. & Evens, M. (1998). "A Computational
Morphology System for Arabic", Proceedings of
COLING-ACL, New Brunswick, NJ.
Censure of the Internet in the Arab countries (2006).
Human Rights Tribune - Geneva 2006 www.humanrights-geneva.info.
Chen, A. & Gey, F. (2002). "Building an Arabic stemmer
for information retrieval".TREC-11 conference.
Darwish, K. (2002). "Building a Shallow Arabic
Morphological Analyzer in One Day". In The ACL-02
Workshop on Computational Approaches to Semitic
Languages, Philadelphia, PA, USA.
De Roeck, A. N. & Al-Fares, W. (2000). "A
morphologically sensitive clustering algorithm for
76
searching".
Information
Processing
and
Management.41 (4), 819-827.
Taghva, K. & Elkoury, R. & Coombs, J. (2005). "Arabic
Stemming without a root dictionary". International
Conference on Information Technology: Coding and
Computing (ITCC'05) - Volume I pp. 152-157.
Thabet, N. (2004). "Stemming the Qur’an" WORKSHOP
ON Computational Approaches to Arabic Script-based
Languages,
University
of
Geneva,
Geneva,
Switzerland, August 28.
Larkey, L. & Ballesteros, L. & Connell, M. (2005).
"Light Stemming for Arabic IR" Arabic Computational
Morphology:
Knowledge-based
and
Empirical
Methods, A.Soudi, A. van en Bosch, and Neumann, G.,
Editors. Kluwer/Springer's series on Text, Speech, and
Language Technology.
Sinane, M. & Rammal, M. & Zreik, K. (2008). "Arabic
documents classification using N-gram", Conference
ICHSL6, Toulouse.
Suleiman Mustafa, H. (2004). "Character contiguity in Ngram based word matching: the case for Arabic text
77