Indexing of Arabic documents automatically based on lexical analysis

International Journal on Natural Language Computing (IJNLC) Vol. 1, No.1, April 2012
1
Indexing of Arabic documents automatically based
on lexical analysis
Abdulrahman Al_Molijy, Ismail Hmeidi, and Izzat Alsmadi
IT faculty, JUST University and Yarmouk University, Jordan
ABSTRACT
The continuous information explosion through the Internet and all information sources makes it
necessary to perform all information processing activities automatically in quick and reliable
manners. In this paper, we proposed and implemented a method to automatically create and Index
for books written in Arabic language. The process depends largely on text summarization and
abstraction processes to collect main topics and statements in the book. The process is developed
in terms of accuracy and performance and results showed that this process can effectively replace
the effort of manually indexing books and document, a process that can be very useful in all
information processing and retrieval applications.
KEYWORDS
Text Summarization, Information Indexing, Text Classification, Text Categorization, Distance
Similarity Measures, Text Stemming And Preprocessing.
1. INTRODUCTION
The Arabic Language
Arabic is a widely spoken language with more than 200 million people, distributed among many
countries in the Middle East region. There are common features and characteristics for the Arabic
language that may not be available in other languages. The Arabic alphabet consists of 28 letters
as follow: ( ‫أ‬
,
‫ب‬
,
‫ت‬
,
‫ث‬
,
‫ج‬
,
‫ح‬
,
‫خ‬
,
‫د‬
,
‫ذ‬
,
‫ر‬
,
‫ز‬
,
‫س‬
,
‫ش‬
,
‫ص‬
,
‫ض‬
,
‫ط‬
,
‫ظ‬
,
‫ع‬
,
‫غ‬
,
‫ف‬
,
‫ق‬
,
‫ك‬
,
‫ل‬
,
‫م‬
,
‫ن‬
,
‫ھـ‬
,
‫و‬
,
‫ي‬ ), in
addition to other forms taken by some characters, such as ( ‫المقصورة‬ ‫أللف‬
)
‫ى‬
(
,
‫المربوطة‬ ‫والتاء‬ )
‫ة‬
(
,
‫والھاء‬
)
‫ه‬
(
‫والھمزة‬ ،
)
‫ء‬ ). Arabic letters contain also points or dots with ten characters that contain one
point each (،‫ن‬ ،‫ف‬ ،‫غ‬ ،‫ظ‬ ،‫ض‬ ،‫ز‬ ،‫ذ‬ ،‫خ‬ ،‫ج‬ ،‫)ب‬, and three letters that contain two points (‫ي‬ ،‫ق‬ ،‫)ت‬, and
two characters that contain three points ( ‫ث‬
,
‫ش‬ ). Different forms of the letters appear according to
their position and one word can consist of more than one part (e.g. a phrase). Letters can have
different forms based on their location in the word. As letters are formed, some of which take one
form, some takes two forms and some other letters take four forms. For example, the letter hamza
(‫)ء‬ takes only one form, and is the only character who takes one form, the letter (waw) (‫و‬, ‫)ـــو‬
takes two forms, the letter eyen (‫)ع‬, which takes one of four forms, (‫ـع‬ ، ‫ـعـ‬ ، ‫عـ‬ ،‫)ع‬. Unlike Latin
languages, Arabic text is read from right to left. Therefore, in any preprocessing to perform the
following steps: Remove punctuation marks, diacritics, non-letters, and stop words.

2
Documents distance similarity methods
In most natural language processing activities, measuring similarity is used for several possible
reasons. For example, it is used to see the most popular words or statements and see
commonalities between those words and statements. It is also used summarize text and eliminate
redundancy in results. Before classification methods, document similarity methods are used to
evaluate the level of similarity between a subject document and the dataset documents. Here are
some examples of distance similarity metrics that are usually used in natural language processing.
Manhattan distance
In Manhattan distance, a rank-order statistic is calculated for two profiles by measuring the
difference in the positions of an N-gram in two different profiles. The general Manhattan
equation is: The class that has the smallest Manhattan distance
(i.e. the closest) is chosen as the class for the document being classified. Example, given two
profiles or documents, P1, P2 consisting of two-grams where P1 = {th, he, ma, an, in, ca, ar}, and
P2 = {ma, he, an, th, in, ca, ar}, then Manhattan distance (P1, P2) = (|0−3| + |1−1| + |2−0| + |3−2|
+ |4−4| + |5−5| + |6−6|) =6. Assume that C1 represents the profile for training class: C1 and
represents the profile for training class C2, and we would like to classify a document with profile
P. After computing the Manhattan distance, if Manhattan (P, C1) < Manhattan (P, C2), the
document will be classified as belonging to class C1.
Dice distance measure
The second similarity measure used is the Dice measure of similarity with the equation: (Pi, Pj)=
2|Pi ∧Pj|/|Pi| + |Pj|. Where |Pi| is the number of elements (N-grams) in profile testing and |Pi| is
the number of elements (N-grams) in profile training. Using the Dice measure, the class with the
largest measure is chosen as the class for the text document being classified. For Example,
assume that the number of N-grams in profile Pi is |Pi| = 20, and the number of N-grams in
profile Pj is |Pj|=10, and that the number of N-grams that are found in both Pi and Pj is |Pi ∧Pj| =
7, then Dice (Pi, Pj) = (2×7)/(10 + 20) = 0.467. Assume that C1 represents the profile for training
class 1 and C2 represents the profile for training class 2, and we would like to classify a
document with profile P. After computing the Dice measure of similarity, if Dice (P, C1) < Dice
(P, C2), the document will be classified as belonging to class C2.
N-gram classification
An N-gram is a subsequence of length N items or characters. An N-gram is like a moving
window over a text, where N is the number of characters in the window. The Size of n-gram can
be 1, 2 and 3 that called unigram, bigram and trigram respectively. If the size of n-gram is 4 or
more, it is called an "n-gram". For example if we have the Arabic word “‫”المسامحة‬, the tri-gram for
this word will be: ( ‫الم‬
,
‫لمس‬
,
‫مسا‬
,
‫سام‬
,
‫امح‬
,
‫محة‬ ).

3
Following are the main tasks in N-gram classification:
First, we have modified N-gram, instead of only splitting the text into tokens consisting of three
letters (tri-gram), we have also implemented algorithm with tokens of four and five letters.
To improve performance, We based classification on the first 100 words rather than the whole
document.
The different options that will be evaluated then depend on preprocessing options, number of
words taken (i.e. 100), number of trained documents (i.e., 1, 3 and 10, 100), and finally the N
number in N-gram ( i.e. 3, 4, or 5). In every cases, the following steps must be hold: ( ., :, ; , ‫ال‬ ,
/, , -, +, ?, <, >, @, $, %, &, *, (, ), !, ~). In every cases, the following steps must be hold:
(1) Split the text into tokens consisting only of letters.
(2) Compute all possible N-grams.
(3) Compute the frequency of occurrences for each N-gram component.
(4) Sort the N-grams according to their frequencies from most frequent to least frequent.
(5) Choose the first 100 repeated words.
(6) This gives us the N-gram profile for a document. For training class documents, the N-gram
profiles were saved to a database.
2. RELATED WORK
Several published papers discussed the usage of different classification algorithms for Arabic text
categorization. A small selection of those papers will be discussed in this section. Rehab Duwairi
has several publications in this subject [5, and 14]. In [5], she discussed a survey of several
published papers in text categorization and evaluated the usage of the classifiers: distance based,
KNN and NB. Results showed that NB is the best classifier out of the three.
Mesleh evaluated the algorithms: SVM and several feature selection methods [6]. His results
showed that CHI square showed the best performance. Baoli et al proposed an improved KNN
algorithm that uses different numbers of neighbours for text categorization [7]. They conducted
preliminary investigation on the Chinese language. Harrag et al proposed using neural networks
for Arabic text classification [8]. They used Singular Value Decomposition (SVD) as a pre-
processor method to neural networks. With several processing stages, they claimed to improve
neural networks to classify Arabic text. They used a total of 453 documents in 14 classes.
Farhoodi and Yari applied machine learning algorithms for Persian text classification [9].
Examples of algorithms they used include SVM and KNN. KNN along with its cosine functions
showed the best ability to detect correct classes for Persian documents. Alsaleem used also SVM
and NB for automatic classification of Arabic text [10]. Author collected a large dataset of 5121
documents of 7 classes. For such large dataset, it will be interesting to see the impact of extending
the classes for more than 7 classes. Based on results, SVM algorithm outperformed NB.
In a similar paper, Bawaneh et al [11] evaluated using SVM and NB for Arabic text classification.
They compared their experiment with several others that were larger in datasets and training. For
the 5 papers they selected, accuracy was between 71-75 to NB and 82-86 for KNN. Such results
conform with most text classification papers in realizing that KNN is the best classifier. Al-

4
Shalabi and Obeidat compared the using of KNN with N-gram classification [12]. N-gram is used
as an alternative for using single words or terms in KNN. The average accuracy in case of using
N-Gram is 0.7353 while with single terms indexing, it is 0.6688. In an earlier paper, Al-Shalabi et
al [13] evaluated the using of KNN for Arabic text classification. However, they didn’t show or
compare results statically with their own or other papers.
An algorithm for Arabic text to extract the word roots is proposed by Al-Shalabi et al. The idea
behind this algorithm is to assign weight and ranks to the letters that form a word based on their
frequent use. The weight is a real number (between 0 to 5, Table 1). Most frequent letters get the
highest weight value or 5. Other high frequent letters get: 3.5, 3, 2, and 1 respectively. The rest of
the Arabic letters get 0 weight. The rank is a real number that depends on the length of the word.
After determining the weights and the ranks for letters in each word, the letters with the smallest
value of multiplying the ranks and the weight represent the root for the word.
Table 1. Assigning Arabic letter weights based on frequency of use
Similar to earlier papers, Thabtah et al evaluated using Chi square, VSM and KNN and NB for
Arabic text classification [15 and 16].
3. GOALS AND APPROACHES
The research and practical contribution in automatic Arabic indexing is minimal. The number of
work in English document indexing is much more in contrast with Arabic document indexing.
The goal is to build an efficient algorithm to perform automatic Arabic indexing with a high level
of accuracy. As discussed earlier, the nature of Arabic text is different from the English text, so
the preprocessing of the Arabic text will be different and a little bit more challenging. Before
applying the indexing process, the documents should be preprocessed. The preprocess phase
includes removing irrelevant text that can be found in the text. This may include: punctuation
marks, diacritics, non-letters, stop words and any words that are not part of the main document
language (e.g. English words when the document is largely written in Arabic). The rest of the
document is retrieved after removing irrelevant text. The number of these words can be large. In
addition ranking techniques are implemented to select the most important terms. After that, to
increase the accuracy of the indexing process, we should extract the roots of the words in order to
reduce the number of the words. There are many methods to extract the root for the word in the
Arabic text (e.g. ref [3]). We will apply a modified version of this method to extract roots of
words in Arabic.

5
Dataset Description
A dataset of Arabic documents is collected from several online Arabic websites such as:
(www.star28.com and www.kutub.info). Each document file size is between 300 to 800 KB and
contains between 1500 to 4000 words. 50 documents from 10 different categories are selected.
The categories of the documents include: Banks, commerce, education, crimes, religion, sport,
tourism, food, sciences, women, are Computer Network and Technology.
Inverted index
Inverted index is an index data structure that saves a mapping from each term in the document to
its locations in a document or a set of documents. The purpose of an inverted index is to allow
fast full text searches, at a cost of increased processing when a document is added to the database.
The inverted file may be the database file itself, rather than its index. It is the most popular data
structure used in document or information retrieval systems such as in search engines.
There are two main variants of inverted indexes: An inverted file index that contains a list of
references to documents for each term. A word level inverted index (or full inverted index or
inverted list) additionally contains the position of each word within the document. The latter form
offers more functionality (e.g. phrase searches). However, it needs more time and space to be
created.
Indexing methods applied
The method that we choose to solve the first part of the system (Automatic Arabic document
indexing problem) is as follow.
The indexing process starts from preprocessing the document by removing irrelevant text such as:
punctuation marks, diacritics, non-letters, stop words and any English word or letter.
Then the system computes the frequency of every term in the document and reorder them
decreasing.
Using ranking techniques will remove all term with highest and lowest frequency.
After that, the system matches between the term and the page number where the term occurs in
the document.
Finally, we add the index to the end of the document.
For the second part of the system (i.e. the inverted index), the process is similar to the first or the
indexing stage with some differences. The developed program first check if the document exists
in the already database of documents. If not, the system starts to preprocess the document and
remove all irrelevant text. Then the system computes the frequency of every term in the document
and reorder them decreasing.

6
Comparison between manual and automatic indexing
In principle, manual indexing is considered as a reference and more accurate than automatic
indexing. This is largely since humans are more capable of dealing with the semantic complexity
of natural language specially in cases where the same word may have different meanings based
on the sentence and paragraph context. Table 1 summarizes the major differences between
manual and automatic document indexing.
Table 1: differences between human and automatic indexing.
Variables Human Indexing Automatic Indexing
Cost expensive per unit indexed
because it is labor-intensive
inexpensive per unit indexed
Time involves more time per unit
indexed
can index large amounts of material
in short amount of time
Extent of
indexable
matter
may be limited to abstract or
summarization of text
routinely based on complete text
Exhaustively tends to be more selective considers most of the words in
indexable material
Specificity uses more generic terminology,
smaller vocabulary
uses very specific terminology,
larger vocabulary
Browsable
displayed
indexes
use multi-term context-providing
headings
limited use of term combinations
Searching
syntax,
display syntax
use wide-range of syntactic
patterns and can adapt quickly to
include newer terminology, as
well as older subject headings
Becoming more sophisticated, and
is selecting, combining,
manipulating, and weighing terms.
Usually limited to key-words in, out
of, or along-side context.
Vocabulary
management
can cross-reference, link
synonyms or like terms, point to
related terms easily
being experimented with
Surrogation not often used by human indexing being used frequently –often as
visual displays, such as icons or
graphs
Evaluation and Analysis
Based on the different criteria described earlier, experiments are conducted to compare results
from automatic indexing relative to manual indexing. Two performance metrics are considered:
Precision and recall. They are calculated according to the following formulas.

7
Figure 1. Prediction performance metrics.
The overall average precision for the 5 documents was 0.998 while recall was 1.00. This simply
means that there was a few values of FP (False Positive), and no value in FN (False Negative).
Accuracy of the results was very high and this is possibly since the size of the dataset is relatively
small.
4. SUMMARY AND CONCLUSION
The large amount of Arabic documents available through the Internet requires the need for an
efficient way to index them for search and retrieval. This paper presents an automatic Arabic
document indexing method. The indexing process starts by implementing preprocessing steps to
remove irrelevant text (e.g. punctuation marks, diacritics, non-letters, etc.). Then the system
computes the frequency of every term in the document and reorders them in a descending order.
A ranking algorithm is then used to remove all terms with highest and lowest frequency. After
that, the system matches between the term and the page number where the term occurs in the
document and automatically adds the index to the end of the document.
The developed program is developed based on accuracy and performance and results showed that
it can effectively replace the human time consuming effort for indexing a large number of
documents or books. Such process can be used and useful for several information retrieval and
processing applications.
5. REFERENCES
[1] Hatem M. Noaman et al, “Naive Bayes Classifier Based Arabic Document Categorization”.
[2] J. He, A. Tan and C. Tan. Comparative Study on Chinese Text Categorization Methods. On the
PRICAI 2000 Workshop on Text and Web Mining, Melbourne, pages 25-31.2000.
[3] Riyad Al-Shalabi et al, “Arabic Text Categorization Using kNN Algorithm,”
[4] Alaa M. El-Halees, “Arabic Text Classification Using Maximum Entropy,” The Islamic University
Journal (Series of Natural Studies and Engineering) Vol. 15, No.1, pp 157-167, 2007
[5] Rehab Duwairi, Arabic text categorization, International Arab journal of information technology, vol.
4, No. 2, April 2007.
[6] Abdelwadood Mesleh, Support Vector Machines based Arabic Language Text Classification System:
Feature Selection Comparative Study, 12th WSEAS Int. Conf. on APPLIED MATHEMATICS, Cairo,
Egypt, December 29-31, 2007.
[7] Li Baoli, Yu Shiwen, and Lu Qin, An Improved k-Nearest Neighbor Algorithm for Text
Categorization, 20th International Conference on Computer Processing of Oriental Languages,
Shenyang, China, 2003.

8
[8] Fouzi Harrag, Abdul Malik Salman Al-Salman, and Mohammed BenMohammed, A Comparative
Study of Neural Networks Architectures on Arabic Text Categorization Using Feature Extraction,
2010 International Conference on Machine and Web Intelligence (ICMWI), 2010
[9] Mojgan Farhoodi, Alireza Yari, Applying Machine Learning Algorithms for Automatic Persian Text
Classification, 2010 6th International Conference on Advanced Information Management and Service
(IMS), 2010.
[10] Saleh Alsaleem, Automated Arabic Text Categorization Using SVM and NB, International Arab
Journal of e-Technology, Vol. 2, No. 2, June 2011.
[11] Mohammed J. Bawaneh, Mahmud S. Alkoffash and Adnan I. Al Rabea, Arabic Text Classification
using K-NN and Naive Bayes, Journal of Computer Science 4 (7): 600-605, 2008.
[12] Riyad Al-Shalabi and Rasha Obeidat, Improving KNN Arabic Text Classification with N-Grams
Based Document Indexing, INFOS2008, March 27-29, 2008 Cairo-Egypt.
[13] Riyad Al-Shalabi, Ghassan Kanaan, and Manaf H. Gharaibeh, (2006). Arabic Text Categorization
Using KNN Algorithm, The 4th International Multiconference on Computer and Information
Technology, CSIT 2006, Amman, Jordan.
[14] Rehab Duwairi, Mohammad Al-Refai, and Natheer Khasawneh, Stemming Versus Light Stemming as
Feature Selection Techniques for Arabic Text Categorization, 4th International Conference on
Innovations in Information Technology, 2007. IIT '07. 2007.
[15] Fadi Thabtah, Mohammad Ali H. Eljinini, AL-Isra, Mannam Zamzeer, and Wa’el Musa Hadi, Naïve
Bayesian Based on Chi Square to Categorize Arabic Data, Communications of the IBIMA, 2009.
[16] Fadi Thabtah, Wa’el Musa Hadi, Gaith Al-shammare, VSMs with K-Nearest Neighbour to Categorise
Arabic Text Data, Proceedings of the World Congress on Engineering and Computer Science 2008
WCECS 2008, October 22 - 24, 2008, San Francisco, USA
[17] Ricardo Baeza-Yates & Berthier Ribeiro-Neto “Modern Information Retrieval” 1st, 1999.
[18] www.wikipedia.org
Sulan ZHANG1, Qing HE, Zheng ZHENG and Zhongzhi SHI, "Automatic Subject Indexing of
Chinese Documents ," Proceeding ofNLP-KE '05, 2005 IEEE.

Indexing of Arabic documents automatically based on lexical analysis

More Related Content

Similar to Indexing of Arabic documents automatically based on lexical analysis (20)

More from kevig (20)

Recently uploaded (20)

Indexing of Arabic documents automatically based on lexical analysis