POLİTEKNİK DERGİSİ
JOURNAL of POLYTECHNIC
ISSN: 1302-0900 (PRINT), ISSN: 2147-9429 (ONLINE)
URL: http://dergipark.gov.tr/politeknik
The corpus based approach to sentiment
analysis in modern standard Arabic and Arabic
dialects: a literature review
Modern standart Arapça ve Arap lehçelerinde
duygu analizine külliyat (korpus) temelli
yaklaşım: bir literatür incelemesi
Yazar(lar) (Author(s)): Anwar ALNAWAS
1
, Nursal ARICI
2
ORCID
1
: https://orcid.org/0000-0001-9181-9377
ORCID
2
: https://orcid.org/0000-0002-4505-1341
Bu makaleye şu şekilde atıfta bulunabilirsiniz(To cite to this article): Alnawas A. and Arıcı N., “The
corpus based approach to sentiment analysis in modern standard Arabic and Arabic dialects: a literature
review”, Politeknik Dergisi, 21(2): 461-470, (2018).
Erişim linki (To link to this article): http://dergipark.gov.tr/politeknik/archive
DOI: 10.2339/politeknik.403975
Politeknik Dergisi, 2018;21(2):461-470 Journal of Polytechnic, 2018;21(2):461-470
461
The Corpus Based Approach to Sentiment Analysis in
Modern Standard Arabic and Arabic Dialects: A
Literature Review
Review Article / Derleme Makalesi
Anwar ALNAWAS
1,3
, Nursal ARICI
2*
1
Graduate School of Natural and Applied Sciences, Computer Engineering, Gazi University, Turkey
2
Department of Computer Engineering, Faculty of Technology, Gazi University, Turkey
3
Nasiriyah Technical Institute, Southern Technical University, Iraq
(Geliş/Received : 13.10.2016; Kabul/Accepted : 21.02.2017)
ABSTRACT
Sentiment Analysis, is the analysis of ideas, emotions, evaluations, values, attitudes and feelings about products, services,
companies, individuals, tasks, events, titles and their characteristics. With the increase in applications on the Internet and social
networks, Sentiment Analysis has taken a considerable place in the field of text mining research and has since been used to explore
the opinions of users about various products or topics discussed over the Internet. When the literature on Sentiment Analysis is
examined, it is seen that the natural language of the Internet information sources that form the basis of the analysis is mostly
English. Developments in the fields of Natural Language Processing and Computational Linguistics have contributed positively to
Sentiment Analysis studies made from natural languages other than English. The purpose of this study is to examine the literature
of Sentiment Analysis conducted in Arabic internet information sources. The literature review includes studies based on the corpus
approach, which is made up of Arabic Internet information sources. Studies are being carried out on the works which constitute
their own corpora for both Modern Standard Arabic and Arabic dialects and on which sentiment analysis is performed.
Keywords: Sentiment Analysis, corpora, feature extraction, Arabic Language, NLP.
Modern Standart Arapça ve Arap Lehçelerinde Duygu
Analizine Külliyat (Korpus) Temelli Yaklaşım: Bir
Literatür İncelemesi
ÖZ
Duygu analizi; kişilerin ürünler, servisler, firmalar, bireyler, görevler, olaylar, başlıklar ve bunların özellikleri üzerine fikirleri,
duyguları, değerlendirmeleri, değer biçmeleri, tutumları ve hislerini analiz edilmesidir. Internet ve sosyal ağlardaki uygulamaların
artmasıyla birlikte, Duygu Analizi (DA), metin madenciliği araştırma alanında dikkate değer bir konuma gelmiş ve o zamandan
beri, kullanıcıların Internet üzerinden tartışılan çeşitli ürünler veya konular hakkındaki görüşlerini keşfetmek için kullanılmaktadır.
Duygu Analizi üzerine yapılan çalışmalar incelendiğinde analize temel oluşturan Internet bilgi kaynakları doğal dilinin çoğunlukla
İngilizce olduğu görülmektedir. Doğal Dil İşleme ve Hesaplamalı Dil Bilim alanlarındaki gelişmeler İngilizce dışındaki doğal
dillerden yapılan Duygu Analizi çalışmalarına olumlu katkıları olmuştur. Bu çalışmanın amacı, Arapça içerikli Internet bilgi
kaynaklarından gerçekleştirilen Duygu Analizi literaturü incelemektir. Literatür incelemesi, Arapça Internet bilgi kaynaklarından
oluşturulan külliyat (corpus) yaklaşımına dayanan çalışmaları kapsamaktadır. Hem Modern Standart Arapça, hem de Arap lehçeleri
için kendi külliyatlarını(corpora) oluşturan ve bu metinler üzerinden duygu analizi yapılan çalışmalar incelenmektedir.
Anahtar Kelimeler: Duygu analizi, külliyat, özellik çıkarma, Arapça, doğal dil işleme
1. INTRODUCTION
The Internet has become an important source of user
generated content such as text, pictures and emoticon in
different sectors including business, health, politics and
education. For that user with every login to Social
Networks generate content in deferent form like
unstructured text, login registration, video streaming and
other forms. The generated content from users is gaining
considerable attention because of its importance in
various sectors like businesses, governmental, political
etc. Understanding what people are thinking and their
views are essential for decision-making on the subject
especially when these comments are combined in one
area and on a single topic.
For the importance of the generated content; Web
Contents Mining (WCM) is being used by organizations
to take advantage of this content. Opinion mining and
Sentiment Analysis (SA) are the sub research area of
WCM appeared to be witnessing increasing research in
the Natural Language Processing (NLP) community.
*Sorumlu Yazar (Corresponding Author)
e-posta : nurs[email protected].tr
Anwar ALNAWAS, Nursal ARICI / POLİTEKNİK DERGİSİ,Politeknik Dergisi,2018;21(2):461-470
462
SA aims to analyze the contents generated by the user,
whether positive or negative feelings about a specific
topic [1, 2]. SA is applied at different levels: document,
sentence and aspect with different techniques. In general
there are two main techniques for SA; lexical and
machine learning approaches. Lexical approach, a
dictionary is prepared to store the polarity values of each
lexicons. For each word of the text present in the
dictionary, polarity score calculating by adding to get an
‘overall polarity score’[3].
Arabic content has grown on the Internet, especially with
the expansion of Social Networks [4]. For example
Twitter has about 11.1 million active users for each
month in Arabic region until March 2017 [5], and
Facebook is the most popular social media platform with
156 million users [6]. In general, textual data available
on the Internet are in both; Modern Standard Arabic
(MSA) and dialects. The dialects of Arabic are very
varied, with five dialects; Gulf (Khaliji), Iraqi, Levantine,
Egyptian and Maghrebi. There is a difference between
the dialects. Some native dialects cannot communicate
with speakers of different dialects, Therefore, MSA is
used for communication and understanding [7, 8].
Standard phonological, morphological and lexical does
not exist in Arabic dialects like MSA [9]. Therefore,
preprocessing is difficult with dialects because it is not
based on fixed rules [10].
In general, many studies on English and some other
languages are widely available, as opposed to Arabic.
The lack of a database of Arabic dialects and studies on
the nature of the acoustic-phonetics are the most
important reasons for natural language processing [11].
A limited number of reviews dealt with Arabic SA. The
study of [12] a comprehensive overview of the latest
updates in the field of SA algorithms and applications.
On the other hand, this work reviewed the Arabic SA
along with the other languages. In [13], reviewed works
on Arabic SA and characteristics of the Arabic language.
The work also reviewed three of SA approaches;
supervised, unsupervised and hybrid.
The proposed paper is a survey on SA of the Arabic
content in various social networks, news or commercial
especially that created its own corpus. As well as the
presentation of the techniques that used to extract
sentiment and corpus building with a review of the most
important challenges to the SA in Arabic language.
This paper is organized as follows: Section 2 provides the
overview of the SA and Arabic language. Section 3
illustrates the works done using corpora for both MSA
and dialects. Section 4 discusses the main issues of
Sentiment Analysis and Arabic language derived from
previous studies. Section 5 concludes the manuscript.
2. BACKGROUND
2.1. Arabic Language
Arabic is an important language, MSA and dialects are
spoken by approximately 422 million people [14]. Arabic
is one of morphology rich language [8].Arabic consists
of 28 letters, including three letters of long vowel and 25
consonants letters. Arabic is written from right to left
alignment. The Arabic letter shape changes according to
its position from the word, unlike English. Arabic also
contains short vowels, or diacritics that change the
meaning of the word such as"
"which means “wrote”
and "
" which means “books”. Arabic is a Semitic
language and contains many dialects depending on
different regions. MSA used in education, news and
official websites.
Arab world describing as an important player in
international politics and the global economy. It is
therefore a focus of attention for multinationals, interest
groups and analysts who want to decipher sentiment on
various issues such as oil prices, stock market movement
and foreign policy. There is a large amount of Arabic
language on the Internet and this quantity needs to
analyze the natural language to extract the required
sentiments [15].
SA for the Arabic language is complex for several
reasons, including characteristics of Arabic language and
lack of SA work for the Arabic language. [4] review the
most important characteristics of the Arabic language
that effect on SA as follow:
-Multiply the root of the Arabic word depending on the
context such as ().
-Punctuation affects the meaning of words, meaning that
two words have the same spelling but different
meanings such as (
) which means grandfather and (
)
which means fined.
-Some linking words can make the sentence carry two
conflicting feelings such as ().
In additional to these points there are different dialects in
each Arabic country or region and this complicates SA
because of different texts that carry feelings.
2.2. Sentiments Analysis
SA is a new field of research uses advanced techniques
to mining texts. Machine learning techniques,
information retrieval and NLP used to handle large
amount of unstructured content that generated by Internet
users, especially social networking sites [16].
SA can be defined as the process of determining the
semantic orientation of the text that holds the opinion.
This semantic orientation can be positive, negative or
natural.SA deals with the text in several ways. Either deal
with a full document as having only one opinion or the
document holds more than one opinion. In general, there
are three levels of SA; document-level, sentence-level
and aspect-based level [17].
•Document-level: at this level, the analysis considers that
the entire document has only one opinion.
•Sentence-level: This level deals with each sentence as
containing one opinion. The polarization of the whole
document depends on the polarization of the sentence.
•Aspect-based level: Sometimes it is called (feature-
based sentiment analysis). A single sentence can contain
more than one aspect and each has its own special
opinion. At this level the desired aspect is determined and
then the polarity of this is set.SA process used two main
THE CORPUS BASED APPROACH TO SENTIMENT ANALYSIS IN MODERN STANDARD … Politeknik Dergisi, 2018; 21 (2) : 461-470
463
approaches: Machine Learning Approach and Linguistics
Based Approach. Figure 1 show the main approaches.
3. CORPORA OF PREVIOUS WORKS
Because of the multitude of dialects in the Arabic
language there’s no corpus containing all dialects.
Create comprehensive corpus requires great abilities that
must be gathered around all dialects in the Arab region.
In general, the texts on the Internet are unorganized and
contain a lot of noise and repetition in the letters. So it
needs to preprocess.
SA of the Arabic language and dialects works, some of
them used the old corpora and the other created new
corpora. In this section we review the works that created
own corpus and how to collect data, source, and method
of preprocess.
3.1. Sentiments Analysis of MSA
In recent years, there has been a marked increase in the
volume of shared data on daily activities of Internet users.
This data contains many information, including user
opinions. The large volume of data requests the
appearance of a new research field called Sentiment
analysis or opinion mining [18].
[19] used the TripAdvisor site to collect 625 review
comments. Comments are categorized manually into five
categories. Comments are categorized into five
categories excellent; very good; middling; weak and
horrible. The final data set contains 250 positive
comments and 250 negative comments. Tokenization and
stemming are applied. They proposed a new
mathematical approach to determine polarity of opinion.
A linear program was designed to calculate the weights
and then calculate the label of each comment. To evaluate
the model, used two terms: Root Mean Square Error
(RMSE) and Average of Margins (AM).
Some researchers has used part of data set previously
collected like in [20], where used large-scale Arabic
Book Reviews Dataset (LABR) that collected by [21] as
a source for Human Annotated Arabic Dataset (HAAD).
HAAD used as annotated corpus for aspect-based
sentiment analysis of Arabic text. To classification of the
comments that selected out of LABR, they used seven
groups of students that studying the course of Natural
Language Processing at Jordan University of Science and
Technology. Each group contains seven students and was
given a task rating of 400 comments. The second phase
of auditing work was conducted by native Arabic speaker
and holds a Ph.D. degree in computer science. The
annotated contains information related to the four tasks:
Aspect Term Extraction (T1), Aspect Term Polarity (T2),
Aspect Category Identification (T3) and Aspect Category
Polarity (T4). Finally, HAAD contains 1513 comments
classified in to positive and negative polarity Aspect
terms. In order to evaluate T1 and T3, the F1 is measured,
where the results were 0.233945 for T1and 0.151815 for
T3.to evaluate T2 and T4, the accuracy of approach is
measured. The accuracy of T2 and T4 was 0.297064 and
0.425743.
[22] used the TripAdvisor site to compile 625 reviews
about hotels. They divided the reviews manually into five
categories: excellent, very good, middling, weak and
horrible. The normalization process was applied,
including the deletion of repeated characters in the word,
delete comments that do not represent an opinion. Some
comments appear in more than one category, in order to
solve this problem, if the comment appears in a category
of 80% it is classified into this category. The proposed
approach used two of ML algorithms: SVM and K-NN.
In first step SVM used to classify in to five classes. To
improve the classification process K-NN used in the
second step to obtain satisfying results. The hybrid
approach showed good results of F-measure up to %97 in
average.
[23] used Facebook as a data source. The case study was
related to the Israel Gaza conflict 2014. Most of the news
on this topic was posted on “breaking news from Gaza”
page. As a first stage, about 10,000 posts were collected.
After the exclusion of the irrelevant posts, remain 2265
posts. Second stage posts were manually annotated to
three classes (positive, negative and neutral) by a group
of three member’s: graduate student and two senior
researchers. BRAT web based annotation tool from [24]
also used. AraNLP tool from [25] used to preprocessing
text that include: tokenization, stemming, segmentation,
part of speech (POS), punctuation and stop word removal
N-Gram. For Name entity recognition (NER) feature
extraction a web service based tool from [26] was used.
The first two tasks T1 and T2 from [20] was applied. The
baseline results for T1 was F1=3762% and for T2 was
Accuracy= 61.47%.
[27] created corpus using data from Twitter and
Facebook. In the beginning 10,000 tweets were collected
from Twitter and classified and examined carefully.
Some problems have been addressed including repetition
in tweets and empty tweets.500 comments were manually
collected from Facebook. Many of these comments have
been excluded because they were written in Latin
characters or contain only emoticons symbols. As a final
result the corpus has 2591 comments and tweets (1073
positive, 1518 negative).Tweets classified using the tool
in [28]. Either Facebook comments classified manually
by the authors of the paper themselves. Using
RapidMiner; Tokenize, Stemming, Filter Stop-words,
and Generate-n-Grams (Terms) operators were applied to
Figure 1: Sentiment classification techniques.
Anwar ALNAWAS, Nursal ARICI / POLİTEKNİK DERGİSİ,Politeknik Dergisi,2018;21(2):461-470
464
the data. The researchers presented an SA of the Arabic
language from the perspective of Machine Learning
(ML).Three supervised methods of classification (Naïve
Bayes, SVM and K-Nearest Neighbor classifiers) were
built using RapidMiner. To split the data into training and
testing sets 10-fold cross validation was used. SVM
achieved the best precision it equals to 75.25%.
[29] collected about 350,000 tweet using Twitter’s
Application Interface (API) by writing PHP script. To
facilitate the classification process, a tool was built to
help the manual classification of tweets. This tool show
to the user each tweet individually and the possibility to
choose the category (positive, negative, neutral and not
applicable).The tweets filtering process was based on
specific criteria like: each tweet contains at least 100
characters, does not contain more than 4 Hashtag, free of
mentions and links and not duplicate or retweets. In the
last stage, more than 25,000 tweets were rated. They
developed RapidMiner extension to match the task of
work. This extension include: Emoticons convertor,
repetitions remover, negation detection, and dialect
(Jordanian dialect) to MSA convertor, Arabizi convertor
and links remover. Three of ML algorithms used to
access proposed framework: NB, SVM and K-NN. The
experimental of work show a good performance of NB
compared to the other.
[21] introduced Large-scale Arabic Book Reviews
(LABR) corpus for SA of the Arabic language. About
220000 review have been downloaded from
“www.goodreads.com”.Approximately 70% of the
review were undesirable because they are not written in
Arabic or not related to Arabic books. After deleting
unwanted contents, the corpus has 63,257 Arabic
reviews. Reviews submitted by users with a rating of 1-
5. The researchers assumed reviews those with ratings 4
or 5 as positive, those with ratings 1 or 2 as negative and
reviews with rating 3 are considered neutral and not
included in the polarity classification. To rating
classification two settings used: a balanced and
unbalanced split of reviews number. Different features:
unigrams, bigrams, and trigrams with/without Term
Frequency–Inverse Document Frequency (TF-IDF)
weighting also used in experiments phase. Scikit-learn
used to implement experiments phase with Python [30].
Table 1: The sentiments analysis for MSA
Studies
Level
Feature
extraction
Data set
source
Classifi
er
Tool
Evaluation
Criteria
Result
[19]
Sentences
Low level
Light
stemming
TripAdvisor
website.
SVM
N/A
RMSE
83.5%
AM
57.6%
[20]
Aspect-
based
N-Grams
LABR
T1;T2;
T3;T4
BRAT
F1-
23.39% T1
15.18% T3
Accuracy
29.7% T2
42.57% T4
[22]
Sentences
Light-
stemming
TripAdvisor
website
SVM+
KNN
N/A
Av. F1
97%
[23]
Aspect-
based
N-Grams,
POS, NER
Facebook
CRF,
J48
AraNLP
F1
37.62% T1
61.47%T2
[27]
Sentences
N-Grams,
Twitter and
Facebook
NB,SV
M K-
NN
RapidMin
er
Macro-
Precision
66.2%NB
75.2%SVM
70.9%KNN
[29]
Sentences
N-Grams,
Twitter
NB,SV
M K-
NN
RapidMin
er
Accuracy
75.4%NB
71.6%SVM
51.5% K-NN
[21]
Sentences
N-Grams,
book readers
social network
www.goodrea
ds.com
MNB,
NB,
SVM
Scikit-
learn
F1
42.6%
MNB
21.1 NB
41.0% SVM
[31]
Sentences
Did not
extracted
PATB,
WTP,
WF
SIMP,
LG
Manual
annotation
Kappa
(k)
82.0% ATB
79% WTP
79.3% WF
[33]
Sentences
N-Grams,
Movies
webpages
NB,
SVM
RapidMin
er
Accuracy
90.6% SVM
89% NB
THE CORPUS BASED APPROACH TO SENTIMENT ANALYSIS IN MODERN STANDARD … Politeknik Dergisi, 2018; 21 (2) : 461-470
465
Multinomial Naive Bayes (MNB), NB (for binary
counts), and SVM used as classifiers. With unbalanced
setting SVM does much better, while in the balanced
setting MNB is slightly better than SVM.
[31] show how to create AWATIF, which represents a
multi-genre corpus of MSA for SA.The dataset that used
were taken from three sources; Penn Arabic TreeBank
(PATB)[32], About 5342 sentences are taken classified
from Wikipedia Talk Pages (WTP) and 2532 of
interrelated conversations taken from 7 Web forum
(WF). To label AWATIF, Identify two types of labeling
guidelines; simple (SIMP) and linguistically-motivated
and genre nuanced (LG). For SIMP, to classify the
sentence to one of three categories (positive, negative, or
natural), they prepared two examples for each of these
species to help annotators. For LG, expose annotators to
a linguistics background and explain the nuances of the
genre to which each data set belongs. With those two
types of guidelines the annotators can classify sentences
under three conditions; GH-LG, GH-SIMP and AMT-
SIMP. GH-LG, Here they use the expertise of students
specializing in linguistics (Referred as Gold Human
(GH)) to classifying sentences using the LG guidelines.
GH-SIMP, Where the GH team works under SIMP
conditions. AMT-SIMP, using Amazon Mechanical Turk
(AMT) as crowd-sourced with SIMP conditions.
[33] collected data from movies web pages in order to
create Opinion Corpus for Arabic (OCA). OCA contains
500 reviews; 250 positive and 250 negative. The process
of OCA creating included the collection of reviews from
several Arabic blog sites and web pages using a simple
bash script for crawling. Manually deleted HTML tags
and special characters, and corrected spelling mistakes.
Tokenizing, removing Arabic stop words, and stemming
and filtering those tokens whose length was less than two
characters also applied using RapidMiner. To evaluate
OCA many of the experiments have been accomplished.
They used cross-validation, Unigram, Bigrams, Trigrams
and TF-IDF, to compare the performance of two of
machine learning algorithms: SVM and NB. Their results
were promising. Table 1 summarizes the sentiments
analysis for MSA which are described above.
3.2. Sentiments Analysis of Arabic Dialects.
Arabic dialects vary according to criteria such as
geography and social class. The following list is only part
of many that covers the main Arabic dialects [34]:
•Gulf Arabic (GLF) includes: of Saudi Arabia Kuwait,
Qatar, Bahrain, Oman, and United Arab Emirates
dialects.
•Iraqi Arabic (IRQ) includes Iraq dialect..
•Sham Arabic (SHM) includes: the of Lebanon, Syria,
Jordan, and Palestine dialects.
•Egyptian Arabic (EGY) includes: Egypt and Sudan
dialects.
•Maghrebi includes: covers the of Morocco, Algeria,
Tunisia, Mauritania and Libya dialects.
•Yemeni dialect.
[35] create two dialects corpora: one for news; and
another for art. The news corpus contains 1000 posts
collected from “Al Arabiyya” news Facebook page. The
corpus of art also contains 1000 posts collected from
“The Voice” page on Facebook. They used three stages
of preprocessing for posts; removing redundancies, time
stamps and Likes. In order to classify the posts manually,
four experts in the Arabic language and dialects were
employed. Five rules applied for classification; Negative,
Positive, Dual, Spam and Neutral. The corpus was
validated, the result of performance was ranging between
73% and 96%.
[36] create a corpus of Mubasher products review using
a small program developed in C # and Twitter’s API.
Over a period of 57 days, 2051 tweets were collected
from Mubasher company Twitter's website in Saudi
Arabia wrote in MSA and local Saudi Arabian dialects.
Two experts rated tweets manually into three classis
(positive, negative and neutral), where labeled positive
tweets as “1”, negative as “-1” and neutral as “0”. After
deleted unreverent tweets, the corpus remains 1331
tweets. In the normalization phase, a set of signs, codes
and Arabic words taken to different forms and English
words were replaced standard Arabic equivalents using
RapidMiner. Tokenization, Removal stop word, Light
stem and Filter token by length. NB and SVM are applied
with two feature selection schemes TF-IDF and BTO
(Binary-Term Occurrence) create the word vector. The
analysis showed good results with SVM.
[37] classified 3700 tweets manually to create their own
corpus. During the classification process, only 1550
tweets appeared related to the specific topic. Users’
names, pictures, hashtags, URLs, emoticons, emoticons,
and all non-Arabic words deleted from Tweets. To
overcome spelling mistakes and to standardize word-
writing formulas in tweets, the normalization process was
applied. Extract feature was based on unigrams, bigrams,
and trigrams. Two classifiers were used to test corpus:
NB and SVM using Weka Suite Software. Researchers
noted that the best results were achieved for both the NB
and SVM classifiers in the unigram language model.
[38] created corpus to predict sentiment in the
commercial sector. The dataset are manually collected
from several web pages like: reviewzat1, jawal1232 and
jumia3. Those data were reviews about five products:
Camera, notebook PC, mobile phone, tablet and
television. Three experts were hired to classify 250
reviews manually, 125 positive and 125 negative were
obtained. Also in this work, they have developed a small
(symbol to word) converter to convert emoticon and
symbol to the corresponding words that match it. Stop
words, special characters, non-Arabic words, and
numbers were deleted in normalization phase. Unigram,
bigrams, trigrams used to extract features. Three of ML
algorithms were used: SVM, NB and K-Nearest
Neighbor (KNN). They realized that the use of SVM and
NB algorithms detecting the polarity of opinions, gives
higher performance than the use of the KNN algorithm.
Anwar ALNAWAS, Nursal ARICI / POLİTEKNİK DERGİSİ,Politeknik Dergisi,2018;21(2):461-470
466
Table 2: The sentiments analysis for dialects Arabic.
Studies
Level
*
Features
extraction
Dialect
Data set
source
Classifier
Tool
Evaluation
Criteria
Result
[35]
Sent
Did not
extracted
Multi
dialects
Facebook
Manual
tagging
Manual
tagging
Accuracy
73-96
[36]
Sent
TF-IDF
and BTO
Saudi
Arabia
n
Twitter
NB, SVM
RapidMiner
Accuracy
83.6%
NB
79.6%
SVM
Precision
78.2%
NB
98.0%
SVM
[37]
Sent
N-Grams
Multi
dialects
Twitter
NB, SVM
Weka
Accuracy
84.2%
NB
84.13%
SVM
F1
83% NB
84%
SVM
[38]
Sent
N-Grams
Multi
dialect
reviewzat1
,
jawal1232,
jumia3
web sites
NB, SVM
Weka
Precision
946%
NB
948%
SVM
F1
939%
NB
948%
SVM
[39]
Sent
N-Grams
Egyptia
n
Dialect
Twitter
ML+SO
Weka
Accuracy
80.9%
F1
80.6%
[40]
Sent
POS
N-Grams
Egyptia
n
dialecta
l
tweets,
product
review,
hotel
reservation
comments
and TV
program
comments
Manual
annotation
N/A
Kappa
97%
[41]
Sent
Unigram,
Bigram
Jordani
an
dialect
Twitter
SVM; NB;
D-Tree;
and KNN
RapidMiner
Accuracy
87.2%
SVM
81.3%
NB
51.45%
D-T
50%
KNN
[42]
Sent
Unigram,
Bigram
Egyptia
n
dialect
Twitter
NB;
SVM
Weka
Accuracy
65% NB
72%
SVM
F1
65% NB
72%
SVM
* Sent= Sentence
THE CORPUS BASED APPROACH TO SENTIMENT ANALYSIS IN MODERN STANDARD … Politeknik Dergisi, 2018; 21 (2) : 461-470
467
[39] used Twitter to create corpus for the Egyptian
dialect. Retrieved 20,000 tweets using Twitter’s API.
Among the retrieved tweets, they used 4800 tweets (1600
positive 1600 negative and 1600 neutral) which every
tweet contained only one opinion. To classify tweets
benefited from two raters. In the initial processing the
images, non-Arabic words, hash tags and URLs were
deleted. To apply preprocessing, the tweets must pass
through the tokenization “is a process of cutting the
document units into small pieces called tokens”.
Stemming and stop-word removal were applied.
Stemming is the process of returning the word to its root
form; and stop-word removal is remove words that have
little meanings such as"" (in), " " (on), " " (of),
etc.. Feature selection was based three different N-Grams
models. The hybrid classifier used a new feature set
combining the ML (NB and SVM) and the semantic
orientation (SO) features. The combination done by
adding the SO score as feature in the feature vector built
using the ML approach; each sentiment word found is
multiplied by the inverse of its SO weight. The results
obtained by proposed hybrid approach showed better
performance than using either ML or SO approaches.
[40] created MIKA corpus consisting of 4000 subjects of
different types of data (tweets, product reviews, hotel
reservation comments and TV program comments)
written in MSA and the Egyptian dialect. Normalization
phase also applied. With the assistance of three Arabic
specialists, the data were classified into three categories;
positive (PO), negative (NG) and neutral (NU) sentiment.
During the classification process, give a power of
sentiment to each topic ranging from 1 to 10 for the
positive, and (-1 to -10) for the negative, and 0 for the
natural.
Several studies have been conducted about SA for
English and two main approaches have been used:
corpus-based and lexicon-based for this purpose. [41]
address both approaches to Arabic starts by create corpus
using Twitter data where the corpus contained 2000
tweets (1000 positive and 1000 negative). The tweets
written in MSA and the Jordanian dialect for different
subjects. The tweets were manually classified by three
experts. Repeated letters and stop-words are removed.
Also normalization for the letters is done by their tool.
They used four supervised methods of classification
using RapidMiner; SVM, NB, D-Tree, and KNN. For
testing and validation, they applied the 5-fold cross
validation. Different stemming techniques tested (root-
stemming, light-stemming and no stemming) on each
classifier. The results show best accuracy with apply
light-stemming for SVM by 87.2%.
[42] use Twitter data to create corpus. Retrieved
approximately 4000 tweets, after extracting tweets
containing only one opinion 1000 tweets were collected;
500 positive and 500 negative. With the help of three
experts, tweets were classified as positive and negative.
Normalization also applied by deleting the user name,
pictures, hash tags, the URLs and all non-Arabic words.
They used Weka Suite Software to extract feature and
apply classifier. Unigrams and bigrams used extract
feature, NB and SVM used as classifier. The results of
the study showed that SVM more accurate than NB,
because NB depends on the possibilities therefore is
suitable for large training dataset. Table 2 summarizes
the sentiments analysis of Arabic dialects that reviewed.
4. DISCUSSION
Social media has become one of the most important of
online activities, and began to attract attention in the
Arabic-speaking countries. Recently, the Arabic
language has begun to raise interest. A lot of the
researches have dealt with the Arabic language within the
field of NLP with current concerns and SAThe most
important issues facing the analysis of feelings for the
Arabic language is the lack of annotated data set. This is
due to several reasons: Arabic has many dialects;
morphological complexity of the Arabic language;
Figure 2: The general framework for corpus based sentiments analysis
Anwar ALNAWAS, Nursal ARICI / POLİTEKNİK DERGİSİ,Politeknik Dergisi,2018;21(2):461-470
468
Arabic dialects differs from MSA phonologically,
morphologically, and lexically.
The dialects included in the previous studies are the
Saudi dialect and the Egyptian dialect with a clear lack of
other dialects. Previous work showed that there were
only two studies [20], [14] that reviewed SA at aspect-
level. While, other studies focused on SA at the sentence
level. From previous studies, we can summarize the basic
steps needed to Arabic SA. Figure 2 show main steps in
general framework for SA.
The general framework for corpus based sentiments
analysis consists of two parts; the first part represents the
establishment of the corpus, model and testing Model.
Corpus creation is done by collecting textual data from
sources such as Facebook or Twitter, then classify these
data manually into the appropriate categories such as
negative and positive. Classified text data is not
normalized (that means writing one word or one letter in
several forms or writing extra characters in the word or
repeating letters within the word) Therefore, the process
of normalization is applied in preprocessing to unify
these different forms.
The preprocessing also contains the delete stop-words
process because these words do not effect on the
classification process. Most of words contains prefix and
suffix and when delete the prefix and suffix we get a large
set of single-root words that improve model.
Using a feature set with low representation of text instead
of full-size text will perform the required task, therefore
input data must be converted to set of features and
carefully extract the features. There are many approaches
to extract features like POS, TF-IDF and N-Gram.
To create SA classification model based on created
corpus, one of the classification algorithms is used or by
using a hybrid method of more than one algorithm.
The created model must be tested to determine its
validity. They are many criteria’s to validate the model
such as F1, accuracy and precision.
The second part of the general framework consists of
collecting the text data from sources, applying
preprocessing and extracting the features as mentioned in
the first part.
5. CONCLUSION
In this work, we reviewed the works related to the
Sentiments Analysis of the Arabic language for both
dialects and Modern Stander Arabic. We provided a
detailed analysis of the methods used and the results
obtained. In the evaluation, it is found that the most
popular methods used to evaluate are NB AND SVM. In
general, the dialects included in the previous studies are
the Saudi dialect and the Egyptian dialect with a clear
lack of other dialects. Also we found that there are
environments that offer support for the Arabic language
such as RapidMiner and Weka. The nature of Arabic was
one of the most important challenges facing Sentiments
Analysis with clear lack of datasets availability.
REFERENCES
[1] Aliane A., Aliane H., Ziane M., and Bensaou N., "A
Genetic Algorithm Feature Selection Based Approach for
Arabic Sentiment Classification", 2016 IEEE/ACS 13th
International Conference of Computer Systems and
Applications (AICCSA), Agadir, Morocco, 1-6, (2016).
[2] Ravi K. and Ravi V., "A Survey on Opinion Mining and
Sentiment Analysis: Tasks, Approaches and
Applications", Knowledge-Based Systems, 89: 14-46,
(2015)
[3] Bhadane C., Dalal H., and Doshi H., "Sentiment
Analysis: Measuring Opinions", Procedia Computer
Science, 45: 808-814, (2015)
[4] Alhumoud S. O., Altuwaijri M. I., Albuhairi T. M., and
Alohaideb W. M., "Survey on Arabic Sentiment Analysis
in Twitter", International Journal of Social, Behavioral,
Educational, Economic, Business and Industrial
Engineering 9: 364-368, (2015)
[5] Internet: WEEDOO, Twitter Arab World – Statistics Feb
2017, 2017, Available: https://weedoo.tech/twitter-arab-
world-statistics-feb-2017/, Accessed: 29 July 2017
[6] Internet: WEEDOO, Facebook Arab World – Statistics
Feb 2017, 2017, Available:
https://weedoo.tech/facebook-arab-world-statistics-feb-
2017/, Accessed: 29 July 2017
[7] Al-Kabi M. N., Gigieh A. H., Alsmadi I. M., Wahsheh H.
A., and Haidar M. M., "Opinion Mining and Analysis for
Arabic Language", International Journal of Advanced
Computer Science and Applications (IJACSA), 5: 181-
195, (2014)
[8] Hamed O. and Zesch T. "The Role of Diacritics in
Designing Lexical Recognition Tests for Arabic", In:
Proceedings of the 3rd International Conference on
Arabic Computational Linguistics, ACLing 2017,
Dubai, United Arab Emirates, 119-128, (2017).
[9] Biskri I., Berrakem F.-Z., and Jebali A. "The Applicative
Combinatory Categorial Analysis of Arabic", In:
Proceedings of the 3rd International Conference on
Arabic Computational Linguistics, ACLing 2017,
Dubai, United Arab Emirates, 199-207, (2017).
[10] Abuata B. and Al-Omari A., "A Rule-Based Stemmer for
Arabic Gulf Dialect", Journal of King Saud University-
Computer and Information Sciences, 27: 104-112,
(2015)
[11] Alshutayri A. and Atwell E., "Exploring Twitter as a
Source of an Arabic Dialect Corpus", International
Journal of Computational Linguistics (IJCL), 8: 37-44,
(2017)
[12] Medhat W., Hassan A., and Korashy H., "Sentiment
Analysis Algorithms and Applications: A Survey", Ain
Shams Engineering Journal, 5: 1093-1113, (2014)
[13] Boudad N., Faizi R., Thami R. O. H., and Chiheb R.,
"Sentiment Analysis in Arabic: A Review of the
Literature", Ain Shams Engineering Journal, (2017)
[14] Internet: UNESCO, Unesco World Arabic Language
Day, 2012, Available:
http://www.unesco.org/new/en/unesco/events/prizes-
and-celebrations/celebrations/international-days/world-
arabic-language-day, Accessed: 23 March 2017
[15] Al-Kabi M. N., Abdulla N. A., and Al-Ayyoub M. "An
Analytical Study of Arabic Sentiments: Maktoob Case
Study", In: Proceedings of the 2013 8th International
THE CORPUS BASED APPROACH TO SENTIMENT ANALYSIS IN MODERN STANDARD … Politeknik Dergisi, 2018; 21 (2) : 461-470
469
Conference for Internet Technology and Secured
Transactions (ICITST), London, UK, 89-94, (2013).
[16] Sharma A. and Dey S. "A Comparative Study of Feature
Selection and Machine Learning Techniques for
Sentiment Analysis", In: Proceedings of the
Proceedings of the 2012 ACM research in applied
computation symposium, San Antonio, Texas,USA, 1-7,
(2012).
[17] Awwad H. and Alpkocak A. "Performance Comparison
of Different Lexicons for Sentiment Analysis in Arabic",
In: Proceedings of the 2016 Third European Network
Intelligence Conference (ENIC), Wrocław, Poland, 127-
133, (2016).
[18] Ibrahim M. A. and Salim N., "Opinion Analysis for
Twitter and Arabic Tweets: A Systematic Literature
Review", Journal of Theoretical & Applied Information
Technology, 56: (2013)
[19] Cherif W., Madani A., and Kissi M. "A New Modeling
Approach for Arabic Opinion Mining Recognition", In:
Proceedings of the 2015 Intelligent Systems and
Computer Vision (ISCV), Fez, Morocco, 1-6, (2015).
[20] Al-Smadi M., Qawasmeh O., Talafha B., and Quwaider
M. "Human Annotated Arabic Dataset of Book Reviews
for Aspect Based Sentiment Analysis", In: Proceedings
of the 2015 3rd International Conference on Future
Internet of Things and Cloud (FiCloud), Rome, Italy,
726-730, (2015).
[21] Aly M. A. and Atiya A. F. "Labr: A Large Scale Arabic
Book Reviews Dataset", In: Proceedings of the 51st
Annual Meeting of the Association for Computational
Linguistics, Sofia, Bulgaria, 494-498, (2013).
[22] Cherif W., Madani A., and Kissi M., "Towards an
Efficient Opinion Measurement in Arabic Comments",
Procedia Computer Science, 73: 122-129, (2015)
[23] AL-Smadi M., Al-Ayyoub M., Al-Sarhan H., and
Jararweh Y. "Using Aspect-Based Sentiment Analysis to
Evaluate Arabic News Affect on Readers", In:
Proceedings of the 2015 IEEE/ACM 8th International
Conference on Utility and Cloud Computing (UCC),
Limassol, Cyprus, 436-441, (2015).
[24] Stenetorp P., Pyysalo S., Topić G., Ohta T., Ananiadou
S., and Tsujii J. i. "Brat: A Web-Based Tool for Nlp-
Assisted Text Annotation", In: Proceedings of the 13th
Conference of the European Chapter of the Association
for Computational Linguistics, Avignon, France, 102-
107, (2012).
[25] Althobaiti M., Kruschwitz U., and Poesio M. "Aranlp: A
Java-Based Library for the Processing of Arabic Text",
In: Proceedings of the Ninth International Conference
on Language Resources and Evaluation, Reykjavik,
Iceland, (2014).
[26] Al-Rfou R., Kulkarni V., Perozzi B., and Skiena S.
"Polyglot-Ner: Massive Multilingual Named Entity
Recognition", In: Proceedings of the Proceedings of the
2015 SIAM International Conference on Data Mining,
British Columbia, Canada, 586-594, (2015).
[27] Duwairi R. and Qarqaz I. "Arabic Sentiment Analysis
Using Supervised Classification", In: Proceedings of the
2014 International Conference on Future Internet of
Things and Cloud (FiCloud), Barcelona, Spain 579-583,
(2014).
[28] Duwairi R., Marji R., Shaban N., and Ershaidat S.,
"Sentiment Analysis", B.S. thesis, Jordan University of
Science and Technology, (2012).
[29] Duwairi R., Marji R., Sha'ban N., and Rushaidat S.
"Sentiment Analysis in Arabic Tweets", In: Proceedings
of the 2014 5th international conference on Information
and communication systems (ICICS), Irbid, Jordan, 1-6,
(2014).
[30] Pedregosa F., Varoquaux G., Gramfort A., Michel V.,
Thirion B., Grisel O., et al., "Scikit-Learn: Machine
Learning in Python", Journal of Machine Learning
Research, 12: 2825-2830, (2011)
[31] Abdul-Mageed M. and Diab M. T. "Awatif: A Multi-
Genre Corpus for Modern Standard Arabic Subjectivity
and Sentiment Analysis", In: Proceedings of the eighth
international conference on Language Resources and
Evaluation (LREC), Istanbul, Turkey, 3907-3914,
(2012).
[32] Maamouri M., Bies A., Buckwalter T., and Mekki W.
"The Penn Arabic Treebank: Building a Large-Scale
Annotated Arabic Corpus", In: Proceedings of the
NEMLAR conference on Arabic language resources
and tools, Cairo, Egypt, 466-467, (2004).
[33] Rushdi‐Saleh M., Martín‐Valdivia M. T., Ureña‐López L.
A., and Perea‐Ortega J. M., "Oca: Opinion Corpus for
Arabic", Journal of the Association for Information
Science and Technology, 62: 2045-2054, (2011)
[34] Biadsy F., Hirschberg J., and Habash N. "Spoken Arabic
Dialect Identification Using Phonotactic Modeling", In:
Proceedings of the Proceedings of the eacl 2009
workshop on computational approaches to semitic
languages, Athens, Greece, 53-61, (2009).
[35] Itani M., Roast C., and Al-Khayatt S. "Corpora for
Sentiment Analysis of Arabic Text in Social Media", In:
Proceedings of the 2017 8th International Conference
on Information and Communication Systems (ICICS),
Irbid, Jordan, 64-69, (2017).
[36] Al-Rubaiee H., Qiu R., and Li D. "Identifying Mubasher
Software Products through Sentiment Analysis of Arabic
Tweets", In: Proceedings of the 2016 International
Conference on Industrial Informatics and Computer
Systems (CIICS), Sharjah, United Arab Emirates, 1-6,
(2016).
[37] Hathlian N. F. B. and Hafezs A. M. "Sentiment-
Subjective Analysis Framework for Arabic Social Media
Posts", In: Proceedings of the Saudi International
Conference on Information Technology (Big Data
Analysis) (KACSTIT), Riyadh, Saudi Arabia, 1-6,
(2016).
[38] Sghaier M. A. and Zrigui M. "Sentiment Analysis for
Arabic E-Commerce Websites", In: Proceedings of the
International Conference on Engineering & MIS
(ICEMIS), Agadir, Morocco, 1-7, (2016).
[39] Shoukry A. and Rafea A. "A Hybrid Approach for
Sentiment Classification of Egyptian Dialect Tweets",
In: Proceedings of the 2015 First International
Conference on Arabic Computational Linguistics
(ACLing), Cairo, Egypt, 78-85, (2015).
[40] Ibrahim H. S., Abdou S. M., and Gheith M. "Mika: A
Tagged Corpus for Modern Standard Arabic and
Colloquial Sentiment Analysis", In: Proceedings of the
2015 IEEE 2nd International Conference on Recent
Anwar ALNAWAS, Nursal ARICI / POLİTEKNİK DERGİSİ,Politeknik Dergisi,2018;21(2):461-470
470
Trends in Information Systems (ReTIS), , Kolkata,
India, 353-358, (2015).
[41] Abdulla N. A., Ahmed N. A., Shehab M. A., and Al-
Ayyoub M. "Arabic Sentiment Analysis: Lexicon-Based
and Corpus-Based", In: Proceedings of the 2013 IEEE
Jordan Conference on Applied Electrical Engineering
and Computing Technologies (AEECT), Amman,
Jordan, 1-6, (2013).
[42] Shoukry A. and Rafea A. "Sentence-Level Arabic
Sentiment Analysis", In: Proceedings of the 2012
International Conference on Collaboration
Technologies and Systems (CTS), Denver, CO, USA,
546-550, (2012).
