A Comparative Study of Root-Based and Stem-Based Approaches For Measuring The Similarity Between Arabic Words For Arabic Text Mining Applications

Advanced Computing: An International Journal ( ACIJ ), Vol.3, No.
6, November 2012
DOI : 10.5121/acij.2012.3607 55

A COMPARATIVE STUDY OF ROOT-BASED AND
STEM-BASED APPROACHES FOR MEASURING THE
SIMILARITY BETWEEN ARABIC WORDS FOR ARABIC
TEXT MINING APPLICATIONS
Hanane FROUD
1
, Abdelmonaim LACHKAR
1
and Said ALAOUI OUATIK
2

1
L.S.I.S, E.N.S.A,University Sidi Mohamed Ben Abdellah (USMBA),Fez, Morocco
[email protected], [email protected]
2
L.I.M, Faculty of Science Dhar EL Mahraz (FSDM), Fez, Morocco
[email protected]

ABSTRACT
Representation of semantic information contained in the words is needed for any Arabic Text Mining
applications. More precisely, the purpose is to better take into account the semantic dependencies
between words expressed by the co-occurrence frequencies of these words. There have been many
proposals to compute similarities between words based on their distributions in contexts. In this paper,
we compare and contrast the effect of two preprocessing techniques applied to Arabic corpus: Root-
based (Stemming), and Stem-based (Light Stemming) approaches for measuring the similarity between
Arabic words with the well known abstractive model -Latent Semantic Analysis (LSA)- with a wide
variety of distance functions and similarity measures, such as the Euclidean Distance, Cosine Similarity,
Jaccard Coefficient, and the Pearson Correlation Coefficient. The obtained results show that, on the one
hand, the variety of the corpus produces more accurate results; on the other hand, the Stem-based
approach outperformed the Root-based one because this latter affects the words meanings.

KEYWORDS
Arabic Language; Latent Semantic Analysis (LSA); Similarity Measures; Root and Light Stemmer.
1. INTRODUCTION
Arabic Documents Representation is challenge and a crucial stage; it may impact positively or
negatively on the accuracy of any Text Mining tasks such as Text Categorization,
Summarization, Documents Clustering, Filtering, and Retrieval purposes. Generally, Arabic
Text Mining applications usually represent documents as Bags-of-Words or Vector Space
Model (VSM) [1][2][3], in which text documents are represented as a set of points in a high
dimensional vector space. However, VSM has four primary limitations which can be grouped
into two problems: the high dimensionality problem and the lack of semantics one. These
limitations are: First, in information retrieval application, a long document gets a low similarity
to a query because the normalized value of the document becomes high. As a result, a long-
length document has little opportunity to match a query. Second, the order of words in a
document is still ignored because of the bag of words assumption. The syntactic structure of a
document is potentially valuable information. Third, keywords in a query have to be exactly
matched with words in documents, and thus the issue of synonymy is not addressed. Fourth and
finally, the issue of polysemy is not addressed because VSM only considers word form.
Therefore, for the Arabic Documents Representation not all features equally represent the
document's semantics; in fact, some of these features may be redundant and add nothing to the
Advanced Computing: An International Journal ( ACIJ ), Vol.3, No.6, November 2012
56

meaning of the document; others might be synonymous and therefore capturing one of them is
enough to enhance the semantic for Documents Representation purposes.
In the other hand, the Arabic Documents Representation may also be impacting by the use of
different text pre-processing approaches, which affect any Text Mining tasks as we have already
concluded in our previous works [7][15].
The main goal of this paper is to compare and contrast the effect of two preprocessing
techniques, that affect the document's semantics, applied to Arabic corpus: Root-based
(Stemming), and Stem-based (Light Stemming) approaches for measuring the semantic between
Arabic words with the well known abstractive model -Latent Semantic Analysis (LSA)- with
different distance functions and similarity measures [15], to overcome the above problems for
Arabic Documents Representation. The LSA model is based on the Singular Value
Decomposition SVD. We used SVD technique to reduce the dimensionality of the vector space
[4] [5], and to build the word representative matrix. This matrix will be used later to quantify
the Arabic words similarity measure. LSA technique [6] [7] is used to quantify the similarity
between Arabic words by their tendency to occur in some contexts than others. The context of a
word [8] consists of a set of tokens distributed on both sides of the word (after and before the
word).
The rest of this paper is organized as follows. The next section introduces the concept of "Latent
Semantic Analysis (LSA)" and its use for measuring similarity between two Arabic words.
Section 3 describes the Stemming techniques for the Arabic Language used in the experiments.
Section 4 discusses the similarity measures and their semantics. Section 5 explains experiment
settings, dataset, results and analysis. In Section 6 we conclude.
2. LATENT SEMANTIC ANALYSIS (LSA)
Latent Semantic Analysis (LSA) is a theory and method for extracting and representing the
contextual-usage meaning of words by statistical computations applied to a large corpus of text.
The underlying idea is that the aggregate of all the word contexts in which a given word does
and does not appear provides a set of mutual constraints that largely determines the similarity of
meaning of words and sets of words to each other.
Latent Semantic Analysis (LSA) model is the automatic procedure proposed by [9] to construct
a vector space. This procedure applies to a vast corpus of texts and includes the three stages
below where the corpus of texts is gradually transformed into semantic vector space of several
hundred dimensions. The text corpus includes two types of separators, paragraphs boundaries
and spaces between words. The paragraph is regarded as the string of characters between two
blanks and the word is the string of characters between two spaces.
The first step of the procedure is to represent the body as a matrix of co-occurrences. The
second is to apply to this matrix a factor analysis called Singular Value Decomposition to get a
space. The last step is to eliminate, among the dimensions of space resulting from the singular
value decomposition, a number of dimensions, regarded as irrelevant.
2.1. Building the Co-occurrence Matrix
For a given corpus, the number of times each word appears in each paragraph is recognized. The
frequencies of co-occurrence between words and paragraphs are calculated. These frequencies
are listed in a matrix. In the Column we find each paragraph, in the row, every word. At the
intersection of a column and row, each cell contains the frequency of occurrences of a word in a
paragraph.

57

Figure 1. Latent Semantic Analysis (LSA)
2.2. The Singular Value Decomposition
The Singular Value Decomposition is a general method of decomposition of a linear matrix in
independents principal components. As a principal components analysis, this method allows to
identify a set of data - here the co-occurrence frequencies - a number of factores uncorrelated
with each other and making each account for the variance of the data set. If n factors account
from the totality of the variance in frequency of co-occurrence, then the data can be represented
in an n-dimensional space, each dimension corresponding to a factor. The table containing the
words in rows and the contexts in the columns form a rectangular matrix , where m is the
number of rows and c the number of columns. This rectangular matrix is decomposed into
three matrices. It is the product, , and :

matrix is a diagonal matrix with n columns and n rows, whose cells contain in the diagonal
"singular values". The word matrix, U, has m lines with n values. The n values in each row are
the coordinates of a vector represented in an n-dimensional space associated with a word
corpus. Each word is represented in an n-dimensional space. After this step, the similarity
between words can then be calculated.
2.3. Reducing the Number of Dimensions
All dimensions emerged from the Singular Value Decomposition is not relevant. The
dimensions associated with the lowest singular values explain only a very small proportion of
(1)
Input Data

Documents
Decomposition
Paragraphs
(Contexts)
Words
The frequencies of
cooccurrence between
words and paragraphs
are calculated
-'- --,--- - ,-
---=- ,-~- ---- -
'-- --'-- -'--- _-
,-
--,---
-'-
---=-
----
,-~-
-'---
Building
the Cooccurrence
Singular Value
Decomposition
Extracting
the Word Matrix
Reducing the
Number of
58

the variance in the original data. If these dimensions were not eliminated, the model would
make errors in estimating the semantic similarity. Since the dimensions are abstract, there are no
criteria for elimination of irrelevant dimensions. Consequently, the number of dimensions
eliminated must be determined empirically. In most applications [6], the number of relevant
factors can go from 30 000 to 300. This means that the vectors pass of 30 000 coordinates in a
space of 30 000 dimensions to 300 coordinates in a space of 300 dimensions.
3. ARABIC TEXT PREPROCESSING
3.1. Arabic Language Structure
The Arabic language is the language of the Holy Quran. It is one of the six official languages of
the United Nations and the mother tongue of approximately 300 million people. It is a Semitic
language with 28 alphabet letters. Its writing orientation is from right-to-left. It can be classified
into three types: Classical Arabic (_=--' ,-'), Modern Standard Arabic (`,-=' ,-') and
Colloquial Arabic dialects (,-'' ,-').
Classical Arabic is fully vowelized and it is the language of the holy Quran. Modern Standard
Arabic is the official language throughout the Arab world. It is used in official documents,
newspapers and magazines, in educational fields and for communication between Arabs of
different nationalities. Colloquial Arabic dialects, on the other hand, are the languages spoken in
the different Arab countries; the spoken forms of Arabic vary widely and each Arab country has
its own dialect.
Modern Standard Arabic has a rich morphology, based on consonantal roots, which depends on
vowel changes and in some cases consonantal insertions and deletions to create inflections and
derivations which make morphological analysis a very complex task [22]. There is no
capitalization in Arabic, which makes it hard to identify proper names, acronyms, and
abbreviations.
3.2. Stemming
Arabic word Stemming is a technique that aim to find the lexical root or stem (Figure 2) for
words in natural language, by removing affixes attached to its root, because an Arabic word can
have a more complicated form with those affixes. An Arabic word can represent a phrase in
English, for example the word '--,--- (tatakrwnana
1
):do you remember us? is
decomposed as follows (Table 1):
Table 1. Arabic Word Decomposition
Postfix Suffix Root Prefix Antefix
'- --
A pronoun
meaning us
Termination of
conjugation
remember
A letter meaning the
tense and the person
of conjugation
Preposition for
asking question

Figure 1.a : Stem Figure 1.b : Root Figure 1.c: Inheritance

Figure 2. An Example of Root/Stem Preprocessing.

1
All Arabic Words in the paper are transliterated using : http://www.russki-mat.net/trans6.html
59

3.3. Root-based versus Stem-based approaches
Arabic stemming algorithms can be classified, according to the desired level of analysis, as
root-based approach (Khoja [10]); stem-based approach (Larkey [11]). In this section, a brief
review on the two stemming approaches for stemming Arabic Text is presented.

Figure 3. Example of Preprocessing with Khoja Stemmer algorithm
Root-Based approach uses morphological analysis to extract the root of a given Arabic word.
Many algorithms have been developed for this approach. Al-Fedaghi and Al-Anzi algorithm
tries to find the root of the word by matching the word with all possible patterns with all
possible affixes attached to it [17]. The algorithm does not remove any prefixes or suffixes. Al-
Shalabi morphology system uses different algorithms to find the roots and patterns [18]. This
algorithm removes the longest possible prefix, and then extracts the root by checking the first
five letters of the word. This algorithm is based on an assumption that the root must appear in
the first five letters of the word. Khoja has developed an algorithm that removes prefixes and
suffixes, all the time checking that its not removing part of the root and then matches the
remaining word against the patterns of the same length to extract the root [10].
The aim of the Stem-Based approach or Light Stemmer approach is not to produce the root of a
given Arabic word, rather is to remove the most frequent suffixes and prefixes. Light stemmer is
mentioned by some authors [19,20,11,21], but till now there is almost no standard algorithm for
Arabic light stemming, all trials in this field were a set of rules to strip off a small set of suffixes
and prefixes, also there is no definite list of these strippable affixes.
In our work, we believe that the preprocessing of Arabic Documents is challenge and crucial
stage. It may impact positively or negatively on the accuracy of any Text Mining tasks;
therefore the choice of the preprocessing approaches will lead by necessity to the improvement
of any Text Mining tasks very greatly.

-----( '-=---- ;-~- ----~ -;-=-
- ,)=-~ -- '-=--- --. ----
---- >=V'- '-= _-=-- --=--
. -- ----- '-=---- ;-~- ----~ -;-=
'-= _-=-- --=- ,)=-~ '-=--
-- >=
;-= ,)= '-= ;-- '-= _-~ -- ;-=
_-= _-=
;-= 2
-- 1
_-~ 1
'-= 2
;-- 1
4
,)= 1
_-= 1
_-= 1
Preprocess
Root-Based
Approach
Term
Weighting
Input Data
Document
Document Processor
and Feature Selection

Removing Stop
Word
Stemming
Term Weighting
Nave Baysian
Classifier
Training
Data
Classified
Document
Text
Mining
Applicati
Results
60

To illustrate this, in Figure 2, we show an opposite example using Khoja stemmer. It produces a
root that is not related to the original word.
For example, the word ('-=--) which means (organizations) is stemmed to ('-=) which
means (he was thirsty) instead of the correct root (=-).
In order to test the effect of the two stemming approaches above on the similarity measures with
the LSA model, we selected tow famous Stemming algorithms: the Morphological Analyzer
from Khoja and Garside [10], and the Light Stemmer developed by Larkey [11].
4. SIMILARITY MEASURES
In this section we discuss the four similarity measures that were tested in [12], and we include
these four measures in our work to effect the measure of the semantic between Arabic words.
4.1. Metric
Not every distance measure is a metric. To qualify as a metric, a measure d must satisfy the
following four conditions. Let x and y be any two objects in a set and
( , ) d x y
be the distance
between x and y.
1. The distance between any two points must be non-negative, that is,
( , ) 0 d x y
.
2. The distance between two objects must be zero if and only if the two objects are
identical, that is,
( , ) 0 d x y =
if and only if
x y =
.
3. Distance must be symmetric, that is, distance from x to y is the same as the distance
from y to x, i.e.
( , ) ( , ) d x y d y x =
.
4. The measure must satisfy the triangle inequality, which is
( , ) ( , ) ( , ) d x z d x y d y z +
.

4.2. Euclidean Distance
Euclidean distance is widely used in clustering problems, including clustering text. It satisfies
all the above four conditions and therefore is a true metric. Given two words
W
a
and
W
b

represented by their vectors
t
a
ur
and
t
b
ur
respectively, the Euclidean distance of the two words is
defined as:
1 2
2
( , ) ( ) ,
, ,
1
m
D t t w w
a t a E b t b
t
=
=
ur ur

where
1, ,
( , ..., )
T
a
a m a
t w w =
r
and
1, ,
( , ..., )
T
b
b m b
t w w =
r
.
4.3. Cosine Similarity
Cosine similarity is one of the most popular similarity measure applied to text documents, such
as in numerous information retrieval applications [13] and clustering too [14]. Given two words
W
a
and
W
b
t
a
ur
and
t
b
ur
respectively their cosine similarity is:
.
( , ) ,
t t
a b
SIM t t
a C b
t t
a b
=
ur ur
ur ur
ur ur

where
1, ,
( , ..., )
T
a
a m a
t w w =
r
and
1, ,
( , ..., )
T
b
b m b
t w w =
r
. The Cosine similarity varies from -1 to 1.

(2)
(3)
61

4.4. Jaccard Coefficient
The Jaccard coefficient, which is sometimes referred to as the Tanimoto coefficient, measures
similarity as the intersection divided by the union of the objects. The formal definition is:
.
( , ) ,
2 2
.
t t
a b
SIM t t
a J b
t t t t
a a b b
=
+
ur ur
ur ur
ur ur ur ur

where
1, ,
( , . . . , )
T
a
a m a
t w w =
r
and
1, ,
( , . . . , )
T
b
b m b
t w w =
r
.
The Jaccard coefficient is a similarity measure and it is 1 when the
t t
a
b
=
ur ur
and 0 when
t
a
ur
and
t
b
ur
are disjoint.
4.5. Pearson Correlation Coefficient
Pearsons correlation coefficient is another measure of the extent to which two vectors are
related. There are different forms of the Pearson correlation coefficient formula. Given two
words
W
a
and
W
b
t
a
ur
and
t
b
ur
respectively, a commonly used form
is:
, 1 ,
( , ) ,
2 2 2 2
, 1 1 ,
m
m w w TF TF
t a a t t b b
SIM t t
a P b
m m
m w TF m w TF
t a a t t t b b

=
=

= =

ur ur

where
, 1
m
T F w
a t a t
=
=
and
1 ,
m
T F w
t b t b
=
=

This is also a similarity measure. However, unlike the other measures, it ranges from -1 to +1
and it is 1 when
t t
a
b
=
ur ur
.
5. EXPERIMENTS RESULTS AND DISCUSSION
Experiments are applied by using two Arabic testing datasets and by applying two schemes of
stemming (Figure 4) : the Larkey's Stemmer developed by [11], and the Khoja's Stemmer [10].
To compute words similarity, we propose to use four schemes of different measures such as the
Euclidean Distance, Cosine Similarity, Jaccard Coefficient, and the Pearson Correlation
Coefficient (Figure 4).
5.1. Dataset
We experimented with two testing datasets both are from the website: http://www.spa.gov.sa/
for the Saudi Press Agency: the first one is a heterogeneous dataset; its composed of 252
documents from different categories (Economics, Politics, and Sports). The second contains 257
documents belonging to one category (politics) [8]. The complete characteristics of the used
corpus are described in Table.2. In the following, for each dataset we experimented with the
above four similarity measures using Larkeys Stemmer (Table.3, Table.4) and using Khojas
Stemmer (Table.5, Table.6).The Euclidean Distance is a distance measure and is bounded
in . For an ideal measure of the similarity between two Arabic words, its Euclidean
distance value 0. In the other hand, the Cosine Similarity, Pearson Coefficient and Jaccard
Coefficient are similarity measures and they are equal to 1 when the words are similar.

(4)
(5)
62

Table 2. Characteristics of the Arabic corpus used in the experiments.
Testing Corpus Characteristics Value
Experiment n1 (DataSet 1)
Number of Documents 252
Size 390 Kbytes
Number of categories 3
Number of Words 31 321
Number of Paragraphs 508
Experiment n2 (DataSet 2)
Number of Documents 257
Size 295 Kbytes
Number of categories 1
Number of Words 23 220
Number of Paragraphs 516

Figure 4. Description of Different Stages of Our Comparative Study

Input Data
Test Corpus
Apply LSA
Model
Preprocessing
Removing Stop Word
: _' ...
Apply the Stemming Approachs

Root-Based
Approach : Khoja
Stem-Based Approach :
Light Stemmer
Select the first 300 values
from representative vectors
Compute Words Similarity

Euclidea
n
Cosine
Similarit
Jaccard
Coefficien
Pearson
Correlatio
Heterogeneous Dataset
3 Categories: Economics, Politics,
and Sports;
Homogeneous Dataset
1 Categorie : Politics;
,-
-
-,---
-
-'-
---=-
,-
-
-,---
-
-'-
---=-
,-
-
-,---
-
-'-
---=-
,-
-
--,---
-'-
---=-
63

5.2. Results
In this work, we propose to compare the different schemes (two stemming schemes and four
different measures) using both Similar Words and Different Words: Similar Words are words
that have the same semantic but they may have different stems or different roots. Different
Words are words that have not the same semantic even they may have the same stem or the
same root.

5.2.1 Results with Larkeys Stemmer
5.2.1. a. Test using Similar Words

In Tab.3, Cosine, Pearson measures are good and the Euclidean distance is not for the pairs like:
(-'---, -) and (,', '), this result shows that the similarity between those Arabic
words is well detected by the LSA model using Larkeys Stemmer. Meanwhile, we can observe,
in the same table, that the LSA model could not reveal the similarity between _- and
', , and -=--- (Cosine, Jaccard and Pearson measures < 0;
Euclidean distance> 0).
5.2.1. b. Test using Different Words
In the Tab.4 we presented the obtained results using the LSA model and Larkeys Stemmer
applied to Different Words, in this table we remark that the used similarity measures performs
worst for the pairs like : ('-',-'-') , which proves that those Arabic words are not similar. But
for (-','', ,,-',') the LSA model affects negatively the similarity between this words, we
obtained a higher Cosine, Jaccard and Pearson measures and a worst Euclidean distances.

5.2.2 Results with Khojas Stemmer
5.2.2. a. Test using Similar Words
The Tab.5 presents the obtained results using the four similarity measures applied to Similar
Words with LSA Model using Khojas Stemmer. In this case, we can observe that the results
show a higher Cosine, Jaccard and Pearson measures, and a worst Euclidean distance. That
prove the similarity between the pairs of words like: (--=, --=) and (,', '). On the
other hand, in the same table, the LSA Model performs worst for the pairs: (-'---, -) and
(-'', ,-') that affects negatively the results obtained in Tab.3.
5.2.2. b. Test using Different Words
In the Tab.6, the higher Cosine, Jaccard and Pearson measures and, in the other side, the worst
Euclidean distance proves that the obtained results for the pairs of Arabic words like: ('--',
,--') are not desirable. In the same table, the bad Cosine, Jaccard and Pearson measures show
good results for the words those are not similar like: ('-', -'-').
The above results show that the obtained results using Larkeys Stemmer outperformed those
obtained using Khojas Stemmer, because this latter affects negatively the obtained results
with LSA Model [15], when we try to measure the similarity between two different Arabic
words that have the same root like: ('--', ,--'), or different root like: (-'---, -),
(-'', ,-'). This is mainly due to the aggressiveness of the Stemming in the sense that it
reduces words to their 3-letters roots.

64

This affects the semantics as several words with different meanings might have the same root
(we can obtain a family of words that can be generated from the same semantic concept from a
single root with different patterns).
The Larkeys Stemmer doesnt produce the root and therefore doesn't affect the semantics of
words; but it maps several words, which have the same meaning to a common syntactical form,
our observation broadly agreed with [16].

Table 3. Similarity Measures for Arabic Words with Larkeys Stemmer using LSA Model:
Using Similar Words.
Experiments Words
Words
Transliterated
English Translation Cosine

Euclidean

Pearson

Jaccard
Experiment
n1
(DataSet 1)
(--=, --=) (md, md) first names of people 0.7023 0.1064 0.7034 0.4801
(~'-~, ~'-~) (sysy, sysy) Political 0.9999 6.575E-14 0.9999 0.9999
(--,-, -,-) (lrqy, lrqy) ( Iraqi man, Iraqi women) 1.000 1.327E-13 0.9999 1.000
(~-,
--~ -'- )
(rfh, wstnkrh) Rejection 0.6302 0.0433 0.6281 0.4501
(,~--, Q---) (llr, wldyn) Religion 0.1262 0.1409 0.1266 0.0613
(Q-V'-
,--~V)
(blmn , wlstqrr) ( Stability, Security) 0.0411 0.1442 0.0414 0.0194
(_~ ,'-) (ww ,wfd) Explain -0.0602 0.1345 -0.0601 -0.02808
(--'~-,
'-'-)
(msdt, nt) Aid 0.02428 0.1212 0.0247 0.0111
(;-V, ;-~-)
(lwm,
lsnwt)
Years 0.1007 0.0932 0.0981 0.0339
(-----, ,-;=--) (ltnmy, ltwyr) Development 0.07793 0.1696 0.07665 0.04052
(~=---, _-,-) (mntb, fryq) Team -0.0261 0.16017 -0.0262 -0.01107
Experiment
n2
(DataSet 2)
(--=, --=) (md, md) first names of people 0.6073 0.1294 0.6069 0.41007
(--,-, -,-) (lrqy, lrqy) ( Iraqi man, Iraqi women) 0.9999 7.741E-14 1.0000 0.9999
(-'---~ ,
~-)
(rfh, wstnkrh) Rejection -0.0137 0.15069 -0.0130 -0.00681
(Q-V'-
,--~V)
(----- ,,-;=--) (ltnmy, ltwyr) Development -0.0335 0.1696 -0.0343 -0.0165

Table 4. Similarity Measures for Arabic Words with Larkeys Stemmer using LSA Model:
Using Different Words.
Experiments Words
Words
Transliterated

Euclide
an

Pearson

Jaccard
Experiment
n1
(DataSet 1)
('-~-, ,--~-) (lsfr, lsfyr) (Ambassador, Embassy) -0.0195 0.17456 -0.02273 -0.00967
('-,-, ~'-,--) (lry, llry) ( Sport, Riyadh) 0.9999 0.0 1.0 0.9999
(~'-,--,Q--~'-,-) (llry,lryyyn) (Sport, athletes) 0.9999 0.0 1.0 0.9999
(---, ~'-,-) (ldwl, lrs) ( Presidency, State) -0.0198 0.1591 -0.0202 -0.0082
(---,----) (ldwl,lbld) (Country ,State) -0.0029 0.1386 -0.0034 -0.0013
(--'~-, ,=--) (msd, lmzr) (Assistant, Massacre) 0.00198 0.1124 0.00112 7.098E-4
(~=----;-->-) (llbwn,lmntb) (Players, Team) 0.00860 0.1257 0.00998 0.004201
('=-=, '-~-) (t, lsbq) (Protest, Race) 0.0165 0.15275 0.02627 0.00815
Experiment
n2
(DataSet 2)
('-~-, ,--~-) (lsfr, lsfyr) (Ambassador, Embassy) -0.05302 0.22078 -0.05136 -0.02518
( --- , ~'-,-) (ldwl, lrs) ( Presidency, State) -0.00415 0.15465 -0.00464 -0.00141
(---,----) (ldwl,lbld) (Country ,State) -0.00221 0.11529
1.077E-
4
-9.433E-4
(--'~-, ,=--) (msd, lmzr) (Assistant, Massacre) 0.04222 0.12945 0.04089 0.01602
(,--~V, )=>- , ) (lstmrr, llhz) (Continuation, Devices) 0.01864 0.15477
0.01800
3
0.00941

65

Though, for the all tables we can see that the variety of the corpus in the experiment n1 gives
more accurate results in both cases (Similar Words or Different Words).
6. CONCLUSION
In this paper, we proposed to compare and contrast the effect of two preprocessing approaches:
Khojas Stemmer (Root-based), and Larkeys Stemmer (Stem-based) for measuring the
similarity between Arabic words using Latent Semantic Analysis (LSA). We experimented with
Experiments Words
Words
Transliterated

Euclidean

Pearson

Jaccard
Experiment n1
(DataSet 1)
(--=, --=) (md, md) first names of people 1.000 1.05E-13 0.999 1.000
(--,-, -,-) (lrqy, lrqy)
( Iraqi man, Iraqi
women)
1.000 0.0 1.0 1.000
(~-, -'---~) (rfh, wstnkrh) Rejection 0.303 0.063 0.304 0.177
(Q-V'-
,--~V)
(_~ ,'-) (ww ,wfd) Explain -0.0237 0.15406 -0.0247 -0.016
(--'~-, '-'-) (msdt, nt) Aid 0.0061 0.1068 0.0051 0.00301
(;-V, ;-~-) (lwm, lsnwt) Years -0.0050 0.1070 -0.0052 -0.0024
(-----, -;=-- , ) (ltnmy, ltwyr) Development 0.01437 0.1323 0.01735 0.0072
(~=---, _-,-) (mntb, fryq) Team 0.0016 0.1109 00021 7.774E-13
Experiment n2
(DataSet 2)
(--=, --=) (md, md) first names of people 1.0 7.14E-14 1.000 0.999
(--,-, -,-) (lrqy, lrqy) ( Iraqi man, Iraqi women) 1.000 0.0 1.0 1.000
(-'---~ , ~-) (rfh, wstnkrh) Rejection -0.086 0.153 -0.089 -0.041
(,~--, Q---) (llr, wldyn) Religion -0.033 0.203 -0.029 -0.01
(Q-V'-
,--~V)
(----- ,,-;=--) (ltnmy, ltwyr) Development 0.00468 0.17135 0.00473 0.00228
Table 5. Similarity Measures for Arabic Words with Khojas Stemmer using LSA Model:
Using Similar Words.

Table 6. Similarity Measures for Arabic Words with Khojas Stemmer using LSA Model:
Using Different Words.

Experiments Words
Words
Transliterated
English
Translation
Cosine

Euclidea
n

Pearson

Jaccard
Experiment n1
(DataSet 1)
('-~-, ,--~-) (lsfr, lsfyr)
(Ambassador,
Embassy)
1.000 1.52E-13 0.999 1.000
('-,-, ~'-,--) (lry, llry) ( Sport, Riyadh) 1.000 0.0 1.0 1.000
(~'-,--,Q--~'-,-) (llry,lryyyn) (Sport, athletes) 1.000 0.0 1.0 1.000
(---, ~'-,-) (ldwl, lrs)
( Presidency,
State)
-0.001 0.089 -0.002 -5.47E-4
(---,----) (ldwl,lbld) (Country ,State) -0.005 0.121 -0.005 -0.002
(--'~-, ,=--) (msd, lmzr)
(Assistant,
Massacre)
0.004 0.172 0.006 0.001
(~=----;-->-) (llbwn,lmntb) (Players, Team) -7.79E-4 0.1093 -0.00127
-3.831E-
4
('=-=, '-~-) (t, lsbq) (Protest, Race) 0.03384 0.11133 0.03238 0.01646
Experiment n2
(DataSet 2)
('-~-, ,--~-) (lsfr, lsfyr)
(Ambassador,
Embassy)
0.999 2.05E-13 0.999 0.999
( --- , ~'-,-) (ldwl, lrs)
( Presidency,
State)
-3.68E-4 0.072 -0.001
-1.840E-
4
(---,----) (ldwl,lbld) (Country ,State) -0.003 0.094 -4.39E-4 -0.001
(--'~-, ,=--) (msd, lmzr)
(Assistant,
Massacre)
0.029 0.224 0.030 0.013
( V ,--~ , ,)=>-) (lstmrr, llhz)
(Continuation,
Devices)
-0.0685 0.2042 -0.0683 -0.0331

66

different schemes of similarity measures. The Root-based approach finds the three-letter roots
for Arabic words without depending on any root or pattern files. The Light Stemming approach
removes the common suffixes and prefixes from the words. The obtained results yield three
conclusions:
1. The Larkeys Stemmer outperforms the Khojas Stemmer because this later affects the
words meanings.
2. Jaccard measure performs bad relatively to the other measures.
3. Cosine and Pearson Correlation measures, and the Euclidean Distance are quite similar for
measuring the similarity between the Arabic words.
We believe that the comparative study presented in this paper will be very important; it may
have a double advantage. First, it may be used generally to support the research in the field any
Arabic Text Mining applications. Second, it will be used precisely to support and guides our
research group to develop correctly our future works.
REFERENCES
[1] Bellot, P. and M. El-Bze. Clustering by means of Unsupervised Decision Trees or Hierarchical and
K-Means-like Algorithm. Proc. of RIAO 2000, 344363.
[2] H.Ghwanmeh (2007). Applying Clustering of Hierarchical K-Means-like Algorithm on Arabic
Language. International Journal of Information Technology IJIT 3, 168172.
[3] G.Salton and M.McGill (1983). Introduction to Modern Information Retrieval. New York:
McGraw-Hill.
[4] H.Schutze and C. Silverstein (1997). Projections for Efficient Document Clustering. In Proceedings
of SIGIR97, 7481.
[5] H.Froud, A.Lachkar, S. Ouatik, and R.Benslimane (2010). Towards a Modern Arabic Information
Retrieval System: Arabic Documents Clustering using Singular Value Decomposition. Journes
Doctorales en Technologies de lInformation et de la Communication JDTIC10,Fs. Morocco.
[6] C.Bellissens, P. and G.Denhire (2004). Deux Modles Vectoriels de la Mmoire Smantique :
Description, Thorie et Derspectives. Le Langage et lHomme.
[7] H.Froud, A.Lachkar, S. Ouatik, and R.Benslimane (2010). Stemming and Similarity Measures for
Arabic Documents Clustering. 5th International Symposium on I/V Communications and Mobile
Networks ISIVC, IEEE Xplore.
[8] C.Barrire and S.Zaki (2001). Rvision dune Hypothse Fausse pour les Mesures de Similarit des
Mots dans les Corpus.
[9] T. LANDAUER and S. DUMAIS (1997). A Solution to Platos Problem: The Latent Semantic
Analysis Theory of the Acquisition, Induction, and Representation of Knowledge. Psychological
Review 104, 211240.
[10] S.Khoja and R.Garside (1999). Stemming Arabic Text. Computing Department, Lancaster
University, Lancaster, http://www.comp.lancs.ac.uk/computing/users/khoja/stemmer.ps.
[11] L.S.Larkey, L.Ballesteros, & M.E.Connell. (2002). Improving Stemming for Arabic Information
Retrieval: Light Stemming and Co-occurrence Analysis. The 25th Annual International Conference
on Research and Development in Information Retrieval (SIGIR 2002), Tampere, Finland, August 11-
15, 2002, (pp. 275-282).
[12] A.Huang, Similarity Measures for Text Document Clustering, NZCSRSC 2008, April 2008,
Christchurch, New Zealand..
[13] B. Yates and R. Neto.Modern Information Retrieval. ADDISON-WESLEY, New York, 1999.
[14] B. Larsen and C. Aone. Fast and Effective Text Mining using Linear-time Document Clustering. In
Proceedings of the Fifth ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining, 1999.
[15] H.Froud, A.Lachkar, and S.Ouatik. (2012). Stemming for Arabic Words Similarity Measures based
on Latent Semantic Analysis Model. The 3rd International Conference on Multimedia Computing
and Systems (ICMCS'12), May 10-12,. Tangier, Morocco, IEEE Xplore.
67

[16] R. Duwairi, M. Al-Refai, and N. Khasawneh, Stemming Versus Light Stemming as Feature
Selection Techniques for Arabic Text Categorization. Innovations in Information Technology, 2007.
IIT '07. 4th International Conference, 446 450.
[17] Al-Fedaghi S. and F. Al-Anzi. A new algorithm to generate Arabic root-pattern forms. In
proceedings of the 11th national Computer Conference and Exhibition. PP 391-400. March 1989.
[18] Al-Shalabi R. and M. Evens. A computational morphology system for Arabic. In Workshop on
Computational Approaches to Semitic Languages, COLING-ACL98. August 1998.
[19] Aljlayl M. and O. Frieder. On Arabic search: improving the retrieval effectiveness via a light
temming approach. In ACM CIKM 2002 International Conference on Information and Knowledge
Management, McLean, VA, USA. PP 340-347. 2002.
[20] Larkey L., and M. E. Connell. Arabic information retrieval at UMass in TREC-10. Proceedings of
TREC 2001, Gaithersburg: NIST. 2001.
[21] Chen A. and F. Gey. Building an Arabic Stemmer for Information Retrieval. In Proceedings of the
11th Text Retrieval Conference (TREC 2002), National Institute of Standards and Technology. 2002.
[22] Abu-Hamdiyyah, Mohammad.2000. The Qur'An: An Introduction

Authors
Miss. Hanane Froud Phd Student in Laboratory of Information Science and Systems,
ECOLE NATIONALE DES SCIENCES APPLIQUES, University Sidi Mohamed
Ben Abdellah (USMBA), Fez, Morocco. She has also presented different papers at
different National and International conferences.

Pr. Abdelmonaime LACHKAR received his PhD degree from the USMBA,
Morocco in 2004 in computer science; He is a Professor in the Head of Computer
Science and Engineering E.N.S.A, University Sidi Mohamed Ben Abdellah
(USMBA), Fez, Morocco. His current research interests include Arabic Text Mining
Applications: Arabic Web Document Clustering and Categorization. Arabic
Information and Retrieval Systems, Arabic Text Summarization, ect , Image
Indexing and Retrieval, 3D ShapeI Indexing and Retrieval in large 3D Objects DataBases, Colour Image
Segmentation, Unsupervised clustering, Cluster Validity Index, on-line and off-line Arabic and Latin
handwritten recognition, and Medical Image Applications.

Pr. Said Alaoui Ouatik is working as Professor in Department of Computer Science, Faculty of
Science Dhar EL Mahraz (FSDM), Fez, Morocco. His research interests include high-dimensional
indexing and content-based retrieval, Arabic Document Categorization. 2D/3D Shapes Indexing and
Retrieval in large 3D Objects DataBase.

A Comparative Study of Root-Based and Stem-Based Approaches For Measuring The Similarity Between Arabic Words For Arabic Text Mining Applications

Uploaded by

Copyright:

Available Formats

A Comparative Study of Root-Based and Stem-Based Approaches For Measuring The Similarity Between Arabic Words For Arabic Text Mining Applications

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Comparative Study of Root-Based and Stem-Based Approaches For Measuring The Similarity Between Arabic Words For Arabic Text Mining Applications

Uploaded by

Copyright:

Available Formats

Advanced Computing: An International Journal ( ACIJ ), Vol.3, No.

You might also like