Pre-indexing Techniques in Arabic Information Retrieval
Souheila Ben Guirat
1,2,4
, Ibrahim Bounhas
2,4
and Yahia Slimani
2,3,4
1
Computer Sciences Department, Prince Sattam Bin Abdulaziz University, K.S.A.
2
Laboratory of Computer Science for Industrial Systems, Carthage University, Tunisia
3
Higher Institute of Multimedia Arts of Manouba (ISAMM), La Manouba University, Tunisia
4
JARIR: Joint Group for Artificial Reasoning and Information Retrieval, Tunisia
(www.jarir.tn)
Keywords: Arabic Information Retrieval, Hybrid Index, Statistical Modeling, Smoothing.
Abstract: Arabic document indexing is yet challenging given the morphological specificities of this language.
Although there has been much effort in the field, developing more efficient indexing approaches is more
and more demanding. One of the most important issues concerns the choice of the indexing units (e.g.
stems, roots, lemmas, etc.) which both enhances retrieval efficiency and optimizes the indexing process. The
question is how to process Arabic texts to retrieve the basic forms which better reflect the meaning of words
and documents? In the literature several indexing units have been compared, while combining multiple
indexes seems to be promising. In our previous works, we showed that hybrid indexes based on stems,
patterns and roots enhances results. However, we need to find the optimal weight of each indexing unit.
Therefore, this paper proposes to contribute in optimizing hybrid indexing. We compare and evaluate four
pre-indexing methods.
1 INTRODUCTION
Indexing process aims to classify documents by
content. Languages with sophisticated grammatical
rules, such as Arabic, require sophisticated indexing
methods.
Alhough there has been a great deal of Arabic
document indexing, there are still indexing problems
that have no been fully solved. One of he most
imporatnt issues is to find the best index term that
faithfully decribes the or user int original word.
2 RELATED WORKS
Identifying terms that discriminate and characterize
the semantics of a document is the main goal of the
statistical indexing (Andersson, 2003). In most
related works in Arabic Information Retrieval (IR),
documents are indexed using stems (Larkey and
Connell, 2001; Aljlayl and Frieder, 2002; Chen and
Gey, 2002) or roots (Al-Kabi et al., 2011; Al-
Shawakfa et al., 2010; Khoja and Garside, 1999).
While Arabic is characterized by its complex
derivational and flectional morphology (Soudani et
al., 2016, Wiem et al., 2015), literature surveys show
that both indexing units may reach better results
according to the experimental settings and the test
collections (Elayeb and Bounhas, 2016). That is,
combining several indexing units is promising and
may reach better results (Ben Guirat et al., 2016).
Consequently, the distinction between different
indexing units (Hadni et al., 2012) is not an essential
question. Anyhow, the most representative index
types are combined in a hybrid indexing approach
(Ben Guirat et al., 2016). However, we need to tune
system parameters by assigning weights to different
indexing units.
Various optimization techniques have been
investigated for other languages including Chinese.
As stated by Shi (2015) the problem of coping with
term dependencies in Chinese is more pervasive than
in most European languages where the bag-of-words
approaches are still considered the state-of-the-art
since they reached good results. In Chinese,
however, phrases are not written as separated words
but as continuous strings of characters. Shi et al.
(2007) showed that combing unigrams with words
and bigrams enhances Chinese IR. The proposed
combining method was based on an empirical deter-
Ben Guirat, S., Bounhas, I. and Slimani, Y.
Pre-indexing Techniques in Arabic Information Retrieval.
DOI: 10.5220/0007393402370246
In Proceedings of the 11th International Conference on Agents and Artificial Intelligence (ICAART 2019), pages 237-246
ISBN: 978-989-758-350-6
Copyright
c
2019 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
237
mination of the linear coefficients of each term.
Some works focused on post-combining
approaches based on merging lists (Kwok, 1997;
Leong and Zhou, 1997) or re-ranking and pseudo-
relevance feedback (Luk and Wong, 2004; Yun et
al., 2005). For example, Leong and Zhou (1997) and
Kwok (1997) merged retrieval lists of words and
bigrams to enhance search effectiveness. Tsang et al.
(1999), Luk et al. (2001) and Chow et al. (2000)
proposed a hybrid indexing approach based on
bigrams and words. They assigned a weight equal to
1.5 for bigrams, while words are weighted according
to their length.
To handle orthographic variants in Japanese,
Kummer et al. (2005) combined words, N-grams,
and Yomi-based indices across different document
collections. From a computational point of view,
they proposed a linear combination of the results of
different retrieval systems and approaches. The
contribution of each system is controlled by a
weight. Relevance feedback is used to gradually
optimize parameters, i.e. the weights of the
individual indexes.
As far as Arabic is concerned, research in index
combination is just starting. Ben Guirat et al. (2016)
combined three indexing units, namely the root, the
stem and the verbed-pattern. For example, the root
of the word "" (“alinkissamat”; the
divisions) is "  " (k s m), its stem is ""
(inkissam; division) and its verbed pattern is ""
(inkassama; was divided). Our goal in this paper is
to enhance hybrid indexing by adopting optimization
in pre-indexing methods which have not been used
in the field of Arabic IR.
3 PROPOSED WEIGHTING
APPROACHES
As presented in the previous section, related works
on Chinese and Japanese languages reveal the
importance of combining more than one indexing
unit. Besides, Ben Guirat et al. (2016) showed the
effectiveness of hybrid indexing compared to single
index-based Arabic IR.
That is, our goal is no longer showing the
evidence of the importance of combining but finding
the best weighting values for each of the indexing
units.
In (Ben Guirat et al. 2016), post-indexing
combining techniques were used. This slightly
enhanced retrieval in Hybrid index IR compared to
basic methods.
In the following, we describe pre-indexing
combination approaches that we propose to further
enhance Arabic IR. We mainly assess linear
combination approaches and smoothing approaches.
3.1 Linear Combination Approaches
In this section, we try to aggregate the weights of
stems (S
i
), roots (R
i
) and verbed patterns (P
i
) to
optimize the search process in the indexing phase.
We propose to combine the frequencies of the three
indexing units with a linear model, as follows:
 
   
   
(1)
where α+ β+ γ=1.
I
j
(W
i
) is the weight of the word Wi in the document
dj and TF
j
(S
i
), (respectively TF
j
(Pi) and TF
j
(R
i
)) is
the normalized frequency of Si (respectively of Pi
and Ri) in document dj.
α, β and γ are three parameters in the interval [0, 1],
which may be varied or estimated in different
manners.
The literature on optimization methods shows a
variety of approaches. Someof the commonly used
optimization approaches (Calandra et al., 2014) are
compared in Table 1.
Gradient descent and Bayesian optimization
reuire more computations which is not suitable for
our combined indexing model that will be tested on
a large amount of data.
Thus, from the list of methods in Table 1, we
chose to implement grid and random search which
seem to be more suitable to our problem because of
their advantages and simplicity required in such
hybrid IR system.
Grid search (Calandra et al., 2014) lies on
running all the combinations of parameters (α and β)
and computing the optimal value of a given IR
metric. In this work, the chosen step size is 0.25.
Using this step size, we aim to cover more values
than previous combining work (Ben Guirat et al.,
2016) which covered only 6 cases (compared to 12
cases in current work). Anyhow this will be refined
in the random search method.
Grid search is costly given the high number of
combinations. Random search tries to reduce the
number of iterations. In this method, the set of
samples is chosen randomly from all the possible
combinations of discrete values of α, β and γ in [0,
1].
ICAART 2019 - 11th International Conference on Agents and Artificial Intelligence
238
Table 1: Comparison of main optimization methods.
Drawbacks
Advantages
Method
-Combinatory
-Grid refinement
gives rise to new
program
iterations(gaps filling
is not applicable)
-Global optimum
-Possible parallelization
Grid
search
-Combinatory in case
of global convergence
-Global optimum
-Possible grid
adjustment
-Less computational
time
-Possible parallelization
-Rapid solution
approximation
Random
search
-Combinatory
-Global optimum
-Probabilistic models
allow to model noisy
observations
Bayesian
optimization
-Requires additional
computations to
gradient evaluations
and parameter
initialization)
-Local optimum.
-Negative influence of
parameter
initialization in global
convergence.
-No possible gap
filling
-Faster convergence
(First order optimizer)
Gradient
descent
Our system implements the following algorithm.
j=1
While (j<threshold)
Aj=RandomSampling (stepsize)
If (performance (Aj)>performance
(Best))
Best=Aj
j++
End While
where Random Sampling (stepsize) is a random
search variant of sampling. It generates a new
position from the hypersphere of a given radius
surrounding the current position.
The Random Search algorithm allows moving
iteratively to better positions in the search space
(Brownlee, 2011) .These positions are sampled from
a hypersphere surrounding the current position.
However, in this algorithm the step size significantly
impacts results. To solve this problem, several
random sampling approaches were proposed in
literature (Brownlee, 2011).
In (Schumer and Steiglitz, 1968), Adaptive Step
Size Random Search (ASSRS) reported the best
results. It is a local search heuristic which changes
dynamically the radius of the hypersphere around
the best solutions to enhance accuracy and to avoid
local optima (Gálvez et al., 2018). It attempts to
heuristically adapt the hypersphere's radius: two new
candidate solutions are generated, one with the
current nominal step size and one with a larger step-
size. The larger step size becomes the new nominal
step size if and only if it leads to a larger
improvement. If for several iterations neither of the
steps leads to an improvement, the nominal step size
is reduced.
In some recent studies, ASSRS is adopted
because of its simplicity and high accuracy (Chen et
al., 2015; Wessing et al., 2017; Gálvez et al., 2018).
In our work, we initialize step size to 0.25 as in grid
search. Then, we apply ASSRS to optimize this
parameter and converge to the best configuration of
indexing units weights.
3.2 Genetic Algorithm with Grid
Search
As in (Cheung et al., 1997), we propose to combine
genetic algorithm (Weise, 2009) with grid search for
better performance and less calculations (Nyarko et
al., 2014; Bergstra and Bengio, 2012).
In this method, we consider only two variables
from (1). The proposed idea is based on problem
composition by optimizing the value of α for each
given value of β; then γ=1-α-β. This is implemented
in the following algorithm.
Do
Generate a set Sβ=( β1, β2, β3… βm)
Generate a set Sα =(α1, α2, α3…αn)
i=1
While (i<n)
j=1
While (j<m)
Optimize (αi)
j++
End While
i++
End While
Until (No significant improvement is
observed)
Pre-indexing Techniques in Arabic Information Retrieval
239
3.3 Smoothing-based Combination
3.3.1 Smoothing Techniques
One of the possible ways to combine the indexing
units is the smoothing technique. It refers to the
adjustment of the maximum likelihood estimator of
a language model so that it will be more accurate
(Zhai and Lafferty, 2001).
Many smoothing algorithms have been proposed
such as additive smoothing (Hazem and Morin,
2013), also called Laplace smoothing. It is one of the
simplest smoothing types but its simplistic
assumption model leads to many drawbacks
including underestimating frequent n-grams and
overestimating unseen ones (Hazem and Morin,
2013). Other alternatives are Good-Turing Estimator
or Katz smoothing extending the intuition of Good-
Turing. Jelineck-Mercer smoothing is also a well-
known smoothing technique. These 4 previously
named techniques all gave good results when tested
for language n-gram modeling (Hazem and Morin,
2013).
Another empirical comparison of smoothing
techniques in language modeling (Chen and
Goodman, 1996), considering multiple set sizes,
performed multiple runs for both bigram and trigram
models. Its results proved again that Katz and
Jelineck-Mercer smoothing perform consistently
well. Church-gale smoothing, which combines Good
Turing with bucketing, outperforms them with
bigrams.
The same work proposes two novel methods:
average count (an instance of Jelinek-Mercer) and
one count method (combining to intuition Makay
and Petro (Chen and Goodman, 1996)). Despite of
the bad performance of one count method, it gives
better results than the other methods.
Besides, Zhai and Lafferty (2001) compared
Jelineck-Mercer, Bayesian smoothing using
Dirichlet Priors (Laplace is a special case for this
technique) and absolute discounting (based on the
similar idea as Jenileck-Mercer). This comparison
aimed to find out the best technique for language
models applied to Ad hoc IR and showed that
Dirichlet Priors is desirable for estimation issues
while Jenileck-Mercer idea suits more query
modeling.
Based on (Federico et al., 2008; Koehn, 2009),
Witten Bell smoothing (Bell et al., 1990) is
considered as well established smoothing technique
as it out-performs many smoothing techniques.
However, many comparison works (Chen and
Goodman, 1996; Federico et al., 2008; Koehn, 2009)
showed that improved Kneser-Ney gives always the
best results an used to perform well even in the
interpolated Kneser-Ney (The, 2006).
3.3.2 Index Weighting as a Smoothing
Problem
We inspired this model from interpolated Kneser–
Ney smoothing. We consider a word represented by
a triplet (
) as a trigram. Our goal is to
compute the weight of each stem S
i
based on its
frequency and the frequencies of its verbed pattern
and its root. We have:




  
   
 
(2)
is given by:

 
In the same manner, we compute

and
.
If
, we consider that
c'
j
(S
i
)
c
j
S
i
=0. This applies
also for

and

.
D is an absolute constant in the interval [0, 1],
for which we may experiment different values. By
default (Stolcke and al., 2011), it is computed as
follows:
   
where
(respectively
) is the total number of
triplets (in our case this is equivalent to the number
of stems) with have exactly one (respectively two)
occurrences in d
j
.
The weight of the stem is obtained by
normalizing
:


(5)
ICAART 2019 - 11th International Conference on Agents and Artificial Intelligence
240
4 EXPERIMENTS
4.1 Test Collection
We tested our approaches in the LDC's standard test
collection ("Arabic Newswire Part 1", catalog number
LDC2001T55). It is composed of 869 megabytes of
news articles taken from "Agence France Presse"
(AFP) Arabic newswire i.e. 383,872 articles dated
from May 13, 1994 through December 20, 2000.
Two versions of TREC topics were developed in
2001 (25 topics) and 2002 (50 topics). Each topic
contains 3 parts, namely a title, a description and
narrative. The later contains further description that
may help the human analyst.
As in some previous works (Ben Guirat et al.,
2016), authors used a modified version of Ghwanmeh
stemmer (Gwanmeh et al., 2009), that appeared to be
more efficient than other stemmers in comparative
studies (Al-Shawakfa et al., 2010). It achieves better
results compared to Khoja and Larkey stemmers (Ben
Guirat et al., 2016). We use PL2 (Poisson estimation
for randomness), which is implemented in Terrier
platform (Ounis et al., 2006), as ranking model.
4.2 Evaluation Protocol
Referred to previous works on LDC2001T55
collection (Soudi et al., 2007), we assess two
scenarios i.e. using only titles and combining titles
with descriptions. We perform four experimental
setups as detailed in table 2.
For each experimental setup, we perform six runs
(cf. Table 3). For measuring search effectiveness,
our comparison is based on 4 metrics, namely Recall,
Precision at 10, Mean Average Precision (MAP) and
R-Precision (Zingla et al., 2018).
Table 2: Experimental setups.
Designation
TREC
version
# topics
Query type
T1
TREC
2001
25
Title
T2
TREC
2002
50
TD1
TREC
2001
25
Title +
description
TD2
TREC
2002
50
Precision is equal to the fraction of documents
retrieved that are relevant to the query, while recall
is the percentage of relevant documents that are
successfully retrieved. To study the ability of our
system to rank documents, we evaluate recall and/or
precision at several positions. For example, Precision
at 10 stands for precision computed for the 10 top
ranked documents. R-precision is equal to precision at
R which is equal to the number of relevant documents
for a given query. In the same perspective, average
precision (AVP) allows to evaluate system
performance by considering precision and recall at
every position in the ranked list. MAP is the average
value of AVP computed for all queries.
4.3 Experimentations Results
Table 3: Compared methods.
Approach
Method
Label
Baselines
Stem based-indexing
S1
Pattern based-indexing
S2
Root based-indexing
S3
Hybrid
indexing
Grid Search based
combination
H1
Random Search based
combination
H2
Genetic-based combination
H3
Kneser-Ney Smoothing
H4
In the following, we start by parameter tuning in grid
search and Kneser-Ney methods (cf. section 1). Then
we compare our approaches using standard IR
metrics.
4.3.1 Grid Search and Kneser-Ney
Parameters Tuning
The goal of this step is to find out the best
parameter configuration that grid search and
Kneser-Ney smoothing technique may reach. We
compare the MAP values of the different values of
parameters (cf. Table 4; Table 5).
Table 4: MAP values in grid search.
Parameter values
T1
T2
TD1
TD2
α
β
γ
0.00
0.25
0.75
0.0136
0.0137
0.0126
0.0120
0.00
0.5
0.50
0.0138
0.0127
0.0130
0.0114
0.00
0.75
0.25
0.0129
0.0152
0.0152
0.0126
0.25
0.00
0.75
0.1601
0.1975
0.1634
0.2355
0.25
0.25
0.50
0.2230
0.2235
0.2214
0.2597
0.25
0.50
0.25
0.2556
0.2269
0.2558
0.2700
0.25
0.75
0.00
0.2421
0.1994
0.2446
0.2477
0.50
0.00
0.50
0.2613
0.2538
0.2654
0.2927
0.50
0.25
0.25
0.2995
0.2728
0.3080
0.3114
0.50
0.50
0.00
0.2924
0.2682
0.3037
0.3079
0.75
0.00
0.25
0.2895
0.2683
0.2966
0.3107
0.75
0.25
0.00
0.3055
0.2764
0.3164
0.3161
Pre-indexing Techniques in Arabic Information Retrieval
241
Table 4 shows that the worst MAP values in grid
search are obtained by the first configurations,
especially when the stem weights are null and give
better results when the stem weights increase. This
fact is due to the nature of the different indexing
units; thus giving important weights to stems which
are naturally the most canonical forms of words,
yields to better MAP values.
Table 4 also shows that all test setups reach the
best results for the same configuration noted GS12
(α=0.75, β=0.25, γ=0). This configuration will be
used in the remaining comparative studies.
Table 5: MAP values variation with D parameter tuning.
MAP
D parameter
T1
T2
TD1
TD2
Default
0.0047
0.0293
0.0023
0.0044
0,1
0.0006
0.0006
0.0004
0.0005
0,2
0.0004
0.0077
0.0003
0.0004
0,3
0.0077
0.0167
0.0002
0.0002
0,4
0.0655
0.0416
0.0001
0.0002
0,5
0.1023
0.1373
0.0670
0.092
0,6
0.0442
0,1056
0.0243
0.0462
0,7
0.0136
0.0648
0.0057
0.0193
0,8
0.005
0.0374
0.0024
0.0093
0,9
0.003
0.024
0.0016
0.056
1
0.0023
0.0171
0.0013
0.0034
In Table 5, we tested the D parameter values
from 0 to 1 using a step= 0.1 but also the default
value (see Eq. 4). The MAP values of all these
possible showed that D=0.5 gives the best MAP
value. So it will be further compared to other
combining techniques in next parts.
4.3.2 MAP Values
In this section, we study the MAP value which is
one of the most important criteria used in IR systems
performance comparison. H1 usually gives high
MAP values. Moreover, it gives the best in T1 and
TD1 setups (cf. Figures 2 and 4). However, S2 gives
the best results in 2002 queries. It may be explained
by the number and the length of the queries as well
as the pool size variations between the two TREC
versions (Voorhees, 2002). Indeed, the average pool
size in 2001 was 164.9 and did not exceed 118.2 in
2002. Kneser-Ney smoothing did not fit our
combining goal and usually gives the worst MAP
values.
Figure 1: MAP values (T1: LDC 2001 titles).
Figure 2: MAP values (T2: LDC 2002 titles).
Figure 3: MAP values (TD1: LDC 2001 titles +
Descriptions).
0
0,05
0,1
0,15
0,2
0,25
0,3
0,35
S1 S2 S3 H1 H2 H3 H4
MAP
0
0,05
0,1
0,15
0,2
0,25
0,3
S1 S2 S3 H1 H2 H3 H4
MAP
0
0,05
0,1
0,15
0,2
0,25
0,3
0,35
S1 S2 S3 H1 H2 H3 H4
MAP
ICAART 2019 - 11th International Conference on Agents and Artificial Intelligence
242
Figure 4: MAP values (TD2: LDC 2002 titles +
Descriptions).
4.3.3 Precision at 10 Values
In this section, we study the values of precision at
10. Figures 5 and 6 show that root and pattern-based
methods have the worst precision rates compared to
the hybrid methods. Furthermore, H3 usually gives
good result that overcomes for all approaches in T1
and TD2.
Besides, all the hybrid approaches (except
Kneser-Ney smoothing method) generally give
better P10 values compared to the baselines except
in T2 (cf. Figure 6) where S1 also gives comparable
MAP value.
Figure 5: Precision at 10 values (T1: LDC 2001 titles).
Figure 6: Precision at 10 values (T2: LDC 2002 titles).
Figure 7: Precision at 10 values (TD1: LDC 2001 titles +
Descriptions).
Figure 8: Precision at 10 values (TD2: LDC 2002 titles +
Descriptions).
0
0,05
0,1
0,15
0,2
0,25
0,3
0,35
S1 S2 S3 H1 H2 H3 H4
MAP
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
S1 S2 S3 H1 H2 H3 H4
P10
0
0,1
0,2
0,3
0,4
0,5
S1 S2 S3 H1 H2 H3 H4
P10
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
S1 S2 S3 H1 H2 H3 H4
P10
0
0,1
0,2
0,3
0,4
0,5
S1 s2 s3 H1 H2 H3 H4
P10
Pre-indexing Techniques in Arabic Information Retrieval
243
4.3.4 R-Precision and Recall Values
Table 6: R-precision and recall Results comparison.
Setup
Approach
R-Precision
Recall
T1
S1
0.3271
0.9471
S2
0.3401
0.9609
S3
0.3325
0.9730
H1
0.3401
0.9602
H2
0.3398
0.9570
H3
0.3212
0.9667
H4
0.1457
0.9228
T2
S1
0.2943
0.9756
S2
0.2955
0.9788
S3
0.2957
0.9801
H1
0.3008
0.9788
H2
0.3017
0.9773
H3
0.2799
0.9801
H4
0.1766
0.9768
TD1
S1
0.3393
0.9876
S2
0.3534
0.9917
S3
0.3443
0.9958
H1
0.3507
0.9917
H2
0.3460
0.9900
H3
0.3269
0.9956
H4
0.0933
0.9454
TD2
S1
0.3372
0.9961
S2
0.3215
0.9971
S3
0.3151
0.9977
H1
0.3390
0.9969
H2
0.3396
0.9962
H3
0.3161
0.9976
H4
0.1362
0.9874
Table 6 compares simple and hybrid approaches
based on two main criteria, namely the R-precision
and recall in all the test setups. For simple indexing
methods, we naturally notice that root-based-
indexing (S3) always reaches the best recall results.
Moreover, we notice the improvement given by
the hybrid approaches compared to basic methods.
Thus, hybrid indexing always reaches better R-
Precision, except in TD1 when S2 gives better R-
Precision results while the chosen smoothing
technique did not improve the IR performance.
Furthermore, using descriptions in queries
usually enhances R-Precision and recall compared to
title-based queries. Actually, descriptions enlarge the
scope of the query by additional terms which may be
synonyms or variants of those existing in the title.
Finally, focusing on the different results of 2001
and 2002 test setups, we note that 2002 queries give
always better results than 2001. This may be
explained by the improvement of number of runs
that have been submitted to TREC 2002 (Soudi,
2007) which enhanced relevance judgment and the
quality of the final collection.
5 CONCLUSION
The contribution of this paper is to use optimization
and smoothing techniques in order to assign weights
to system parameters in the pre-indexing stage.
To get a closer representation of the importance
of each indexing unit in representing word meaning,
we used the LDC's standard test collection which is
covering more vocabulary than ZAD collection used
in previous combining works (Ben Guirat et al.,
2016). This test collection also contains more
concise queries, which include detailed descriptions
that gave us the opportunity to study the effect of
query length in retrieval effectiveness. This allowed
us to obtain better results and study the specificities
of each approach/configuration.
The presented results clearly show that our
proposed approaches which combine different
indexing units usually outperform simple indexing.
Especially, the grid search usually gives the best
performance with its optimal weights values
(α=0.75, β=0.25, γ=0). However, the variety of the
number of queries and their length shows variations
between the 4 setup results.
Further, we would like to assess other combining
methods like other smoothing techniques (Zhai and
Lafferty, 2001), since Kneser-Ney smoothing did
not improve our IR system. Regression approaches
(Lamprier et al., 2007) could also be used to
estimate the units' weights values. We also plan to
integrate other stemming tools to process texts. For
instance, stem-based indexing with FARASA (El
Mahdaouy et al., 2018) and lemma-based indexing
with MADAMIRA enhanced Arabic IR (Soudani et
al., 2018).
REFERENCES
Abdelhadi Soudi, Antal van den Bosch, and Gunter
Neumann, 2007. “Arabic Computational Morphology:
Knowledge-based and Empirical Methods”, Springer
Publishing Company, Incorporated.
Abdelkader El Mahdaouy, Said El Alaoui Ouatik, and
Éric Gaussier, 2018. “Improving Arabic information
retrieval using word embedding similarities”. I. J.
Speech Technology 21(1): 121-136.
Aitao Chen, and Frederic Gey, 2002. “Building an Arabic
stemmer for information retrieval”, Proceedings of the
Text Retrieval Conference TREC-11, pp. 631-639.
Akemi Gálvez, Andres Iglesias, Iztok Fister, Iztok Fister
Jr., Eneko Osaba, and Javier Del Ser, 2018. “Memetic
Modified Cuckoo Search Algorithm with ASSRS for
the SSCF Problem in Self-Similar Fractal Image
Reconstruction”. HAIS 2018: 658-670.
ICAART 2019 - 11th International Conference on Agents and Artificial Intelligence
244
Amir Hazem and Emmanuel Morin, 2013. “A Comparison
of Smoothing Techniques for Bilingual Lexicon
Extraction from Comparable Corpora”, Proceedings of
the Workshop on Building and Using Comparable
Corpora (BUCC’13).
Andreas Stolcke, Jing Zheng, Wen Wang and Victor
Abrash, 2011. “SRILM at sixteen: Update and
outlook”. In Proceedings IEEE Automatic Speech
Recognition and Understanding Workshop.
Bernard K. S. Cheung, André Langevin, and Hugues
Delmaire, 1997. “Coupling genetic algorithm with a
grid search method to solve mixed integer nonlinear
programming problems”, Computers and Mathematics
with Applications. vol. 34, no. 12,pp. 13-23.
Bilel Elayeb, and Ibrahim Bounhas, 2016. “Arabic Cross-
Language Information Retrieval: A Review”. ACM
Trans. Asian & Low-Resource Lang. Inf. Process.
15(3): 18:1-18:44.
Chengxiang Zhai and John Lafferty, 2001. “A study of
smoothing methods for language models applied to Ad
Hoc information retrieval”, Proceedings of the 24th
annual international ACM SIGIR conference on
Research and development in information retrieval,
p.334-342, New Orleans, Louisiana, USA.
Ellen M. Voorhees, 2002. “Overview of TREC 2002”. In
Proceedings of the 11th Text Retrieval Conference
(TREC 2002), NIST Special Publication 500-251,
National Institute of Standards and Technology.
Emad Al-Shawakfa, Amer Al-Badarneh, Safwan
Shatnawi, Khaleel Al-Rabab’ah, and Basel Bani-
Ismail, 2010. “A comparison study of some Arabic
root finding algorithms”, Journal of the American
Society for Information Science and Technology, vol.
6, no. 5, pp. 1015–1024..
Emmanuel K. Nyarko, Robert. Cupec, and Damir Filko,
2014. “A Comparison of Several Heuristic Algorithms
for Solving High Dimensional Optimization
Problems”, International journal of electrical and
computer engineering systems, Vol.5. No.1, pp. 1-8.
Iadh Ounis, Gianni Amati, Vassilis Plachouras, Ben He,
Craig Macdonald, and Christina Lioma, 2006.
“Terrier: A High Performance and Scalable
Information Retrieval Platform”. In Proceedings of
Open Source Information Retrieval (OSIR) Workshop,
Seattle, USA.
James Bergstra, and Yoshua Bengio, 2012. “Random
search for hyper-parameter optimization”, The
Journal of Machine Learning Research, vol. 13, no. 1,
pp.281-305.
Jason Brownlee, 2011. “Clever Algorithms: Nature-
inspired Programming Recipes”, Lulu.com.
Ken C.W.Chow, Robert W. P Luk, Luk and K.am-Fai
Wong, Kuilam Kwok, 2000. “Hybrid term indexing
for different IR models”, Proceedings of the 5th
International Workshop Information Retrieval with
Asian Languages, pp.49-54.
Kuilam Kwok, “Comparing Representations in Chinese
Information Retrieval”, 1997. Proceedings of 20th
annual international ACM SIGIR conference on
Research and development in information retrieval,
pp. 34-41, Philadelphia, PA, USA .
Leah Larkey, and Margaret E. Connell, 2001. “Arabic
information retrieval at UMass in TREC-10”,
Proceedings of Text Retrieval conference (TREC)
2001, Gaithersburg, USA.
Linda Andersson, 2003. “Performance of Two Statistical
Indexing Methods, with and without Compound-word
Analysis”, English summary of a bachelor thesis,
Stockholm University, Sweden, available
online:www.nada.kth.se/kurser/kth/2D1418/uppsatser
03/LindaAndersson_compound.pdf.
Lixin Shi, Jian Y. Nie, Jing Bai, 2007. “Comparing
Different Units for Query Translation in Chinese
Cross-Language Information Retrieval”, Proceedings
of the 2nd International Conference on Scalable
Information Systems (Infoscale), Suzhou, China.
Lixin Shi, 2015. “Relating dependent terms in information
retrieval”, Ph.D thesis, Monreal University, Canada.
Ma Schumer, and Kenneth Steiglitz, 1968. “Adaptive
step size random search”. IEEE Transactions on
Automatic Control, 13(3), 270-276.
Marcello Federico, Nicola Bertoldi and Mauro Cettolo,
2008. “IRSTLM: An open source toolkit for handling
large scale language models”, INTERSPEECH 2008,
9th Annual Conference of the International Speech
Communication Association, Brisbane, Australia.
Mariem Amina Zingla, Chiraz Latiri, Philippe Mulhe,
Catherine Berrut, and Yahia Slimani, 2018. “Hybrid
query expansion model for text and microblog
information retrieval”, Information Retrieval Journal,
pp 1–31, Inf Retrieval J, Springer Netherlands.
Meryeme Hadni, Abdelmonaime Lachkar, and Said O.
Alaoui, 2012. “A New and Efficient Stemming
Technique for Arabic Text Categorization”, ICMCS
2012 International Conference on Multimedia
Computing and Systems, pp. 791 - 796, Tangier,
Morocco.
Mohamed Aljlayl and Ophir Frieder, 2002. “On arabic
search: improving the retrieval effectiveness via a light
stemming approach”, the 11th Conference on
Information and Knowledge Management, McLean,
Virginia, USA, pp. 340-347.
Mohammed Al-Kabi, Qasem Al-Radaideh and Khalid
Akawi, 2011. “Benchmarking and assessing the
performance of Arabic Stemmers”, Journal of
Information Science, vol. 37, no. 2, pp. 1–12.
M-K. Leong, and H. Zhou, 1997. “Preliminary qualitative
analysis of segmented vs bigram indexing in Chinese”,
Proceedings of the Sixth Text Retrieval Conference
(TREC-6), Gaithersburg, Maryland, USA.
Nadia Soudani, Ibrahim Bounhas, and Sawssen Ben Babis,
2018. “Ambiguity Aware Arabic Document Indexing
and Query Expansion: A Morphological Knowledge
Learning-Based Approach”. FLAIRS Conference
2018: 230-23.
Nadia Soudani, Ibrahim Bounhas,and Yahia Slimani,
2016. “Semantic Information Retrieval: A
Comparative Experimental Study of NLP Tools and
Pre-indexing Techniques in Arabic Information Retrieval
245
Language Resources for Arabic”. ICTAI 2016: 879-
887.
Nina Kummer, Christa Womser-Hacker, Noriko Kando,
2005. “Handling Orthographic Varieties in Japanese
Information Retrieval: Fusion of Word-, N-gram, and
Yomi-Based Indices across Different Document
Collections”, Proceedings of the Second Asia
conference on Asia Information Retrieval Technology,
pp. 666-672, Jeju Island, Korea.
Philipp Koehn, “Statistical Machine Translation”,
Cambridge University Press; 1 edition, 17 December
2009.
Robert W. P. Luk, Kam-Fai Wong, and Kuilam Kwok,
2001. “Hybrid term indexing: an evaluation”,
Proceedings of the 2nd NTCIR Workshop on Research
in Chinese & Japanese Text Retrieval and Text
Summarization, pp.130-136, Tokyo Japan.
Robert W.P. Luk and Kam-Fai Wong, 2004. “Pseudo
Relevance Feedback and Title Re-ranking for Chinese
Information Retrieval”, Proceedings of the Fourth
NTCIR Workshop on Research in Information Access
Technologies, Information Retrieval, Question
Answering and Summarization, Tokyo, Japan.
Roberto Calandra, Nakul Gopalan, André Seyfarth, Jan
Peters, and Mark P. Deisenroth, 2014. “Bayesian gait
optimization for bipedal locomotion”, in " Learning
and Intelligent Optimization", proceedings of 8th
International Conference, Lion 8, Gainesville, FL,
USA, February 16-21, 2014 (Revised Selected Papers)
Lecture Notes in Computer Science, pp. 274-290.
Sameh Ghwanmeh, Saif Addeen Rabab’ah, Riyad Al-
Shalabi, and Ghassen Kanaan, 2009. “Enhanced
Algorithm for Extracting the Root of Arabic Words”.
Sixth International Conference on Computer
Graphics, Imaging and Visualization, pp. 388–391,
Tianjin, China..
Shereen Khoja, and Roger Garside, 1999. “Stemming
Arabic Text. Technical report”, Computing
department, Lancaster University, U.K.
Simon Wessing, Rosa Pink, Kai Brandenbusch, and
Gunter Rudolph, 2017. “Toward Step-Size Adaptation
in Evolutionary Multiobjective Optimization”. EMO
2017: 670-684.
Souheila Ben Guirat, Ibrahim Bounhas, and Yahia
Slimani, 2016. “Combining Indexing Units for Arabic
Information Retrieval”. International Journal of
Software Innovation (IJSI), vol. 4, no. 4, pp. 1-14.
Souheila Ben Guirat, Ibrahim Bounhas, and Yahia
Slimani, 2016. “A Hybrid Model for Arabic Document
Indexing”, in 17th IEEE/ACIS International
Conference on Software Engineering, Artificial
Intelligence, Networking and Parallel/Distributed
Computing (SNPD 2016), May 30 - June 1, 2016,
Shanghai China.
Souheila Ben Guirat, Ibrahim Bounhas, and Yahia
Slimani, 2018. “Enhancing Hybrid Indexing for
Arabic Information Retrieval”, 32nd International
Symposium, ISCIS 2018, Held at the 24th IFIP World
Computer Congress, WCC 2018, Poznan, Poland.
Stanley F. Chen and Joshua Goodman, 1996. “An
empirical study of smoothing techniques for language
modeling”, Proceedings of the 34th annual meeting
on Association for Computational Linguistics, p.310-
318, Santa Cruz, California.
Sylvain Lamprier, Tassadit Amghar, B. Levrat, and
Frederic Saubion, 2007. “Document Length
Normalization by Statistical Regression”, 19th IEEE
International Conference on Tools with Artificial
Intelligence(ICTAI 2007), Patras, Greece.
T.F. Tsang, R.W.P. Luk, and K.F. Wong, 1999. “Hybrid
term indexing using words and bigrams”, Proceedings
of the fourth International Workshop on Informatoon
Retrieval with Asian Languages (IRAL), pp. , 112-117,
Academia Sinica, Taiwan.
Thomas Weise, 2009. “Global Optimization Algorithms -
Theory and Application”, Self-Published, second
edition.
Timothy C. Bell, John G. Cleary and Ian H. Witten, 1990.
“Text Compression”. Prentice Hall, Englewood Cliffs,
N.J,
Wiem Lahbib, Ibrahim Bounhas, and Yahia Slimani.
2015. “Arabic Terminology Extraction and
Enrichment Based on Domain-Specific Text Mining”.
ICTAI 2015: 340-347.
Yee Whye The, 2006. “A Bayesian interpretation of
interpolated Kneser-Ney”, Technical Report TRA2/06,
School of Computing, National University of
Singapore.
Yu Chen, Jian Cao, and Pinglei Guo, 2015. “CPU Load
Prediction Based on a Multidimensional Spatial
Voting Model”. DSDIS 2015: 97-102.
Yun Xiao, Robert W.P. Luk, Kam-Fai. Wong, Kuilam
Kwok, 2005. “Some Experiments with Blind
Feedback and Re-ranking for Chinese Information
Retrieval”, Proceedings of the Fifth NTCIR Workshop
Meeting on Evaluation of Information Access
Technologies: Information Retrieval, Question
Answering and Cross-Lingual Information Access,
Tokyo, Japan.
ICAART 2019 - 11th International Conference on Agents and Artificial Intelligence
246