Deep Learning For Sentiment Analysis of Tunisian D
Deep Learning For Sentiment Analysis of Tunisian D
net/publication/349343533
CITATIONS READS
0 218
3 authors:
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Hamdi Jamila on 01 March 2021.
University of Sfax,
ANLP Research group, MIRACL Lab.,
Tunisia
Abstract. Automatic sentiment analysis has become In this respect, our work is part of the automatic
one of the fastest growing research areas in the analysis of Internet users’ comments that are
Natural Language Processing (NLP) field. Despite its posted on the official pages of supermarkets in
importance, this is the first work towards sentiment Tunisia on Facebook social networks. To do this,
analysis at both aspect and sentence levels for the We have gathered comments from the official
Tunisian Dialect in the field of Tunisian supermarkets.
Facebook pages of Tunisian supermarkets.
Therefore, we experimentally evaluate, in this paper,
three deep learning methods, namely convolution neural To conclude, the main objective of this research
networks (CNN), long short-term memory (LSTM), is to propose an automatic sentiment analysis of
and bi-directional long-short-term-memory (Bi-LSTM). Internet users’ comments that are posted on the
Both LSTM and Bi-LSTM constitute two major types official pages of Tunisian supermarkets and on
of Recurrent Neural Networks (RNN). Towards this Facebook social networks. We began by studying
end, we gathered a corpus containing comments
the issues related to the term ‘opinion’ and the
posted on the official Facebook pages of Tunisian
supermarkets. To conduct our experiments, this
existing solutions. At the end of this study, we
corpus was annotated on the basis of five criteria proposed a method for the analysis of feelings.
(very positive/positive/neutral/negative/very negative) The primary contributions of this paper are
and other twenty categories of aspects. In this as follows:
evaluation, we show that the gathered features can lead
to very encouraging performances through the use of
CNN and Bi-LSTM neural networks. — We gather comments from the official Face-
book pages of Tunisian supermarkets, namely
Keywords. Sentiment Analysis; Tunisian Dialect; Social Aziza, Carrefour, Magasin Genéral, Géant
networks; Aspect-based Sentiment Analysis; Sentence- and Monoprix. This corpus is made up of
Based sentiment analysis; Big data; CNN; RNN. comments written in the Tunisian Dialect by
taking into account two main scripts; Latin
1 Introduction script and Arabic script.
The remaining of this paper is structured as Three main classifiers were adopted for the clas-
follows. Section 2 surveys the literature on sification task, namely Support vector machines,
sentiment analysis. In Section 3, we deal with the Bernoulli naı̈ve Bayes and Multilayer perceptron.
Tunisian Dialect dataset that were collected and Their obtained error rates were 0.23, 0.22,
used in our experiments. Section 4 is dedicated and 0.42 for support vector machine, multilayer
to a detailed presentation of our proposed method perceptron and Bernoulli naı̈ve Bayes, respectively.
for opinion analysis in the Tunisian dialect. In Deep learning is a crucial part of machine
this context, we propose two different approaches: learning that refers to the deep neural network
The first allows to classify the collected comments suggested by G.E. Hinton [56]. It includes five
into five sentiment classes at the sentence main networks or architectures: CNN (convolu-
level, while the second aims to introduce the tional neural networks), RNN (recursive neural
aspect-based sentiment analysis of the Tunisian networks), RNN (recurrent neural networks),
dialect by implementing two models, namely the DBN (deep belief networks), and DNN (neural
aspect category model and the sentiment model. networks deep).
Section 5 provides a discussion that demonstrates
the efficiency and accuracy of both RNN and
2.2 Knowledge-Based Approach
CNN-based features. We finally draw some
conclusions and future work directions in Section 6. The knowledge-based approach also known
as lexicon-based approach consists in building
2 The Main Approaches of Sentiment lexicons of classified words. In this respect, [55]
relied on a lexicon-based approach to be able
Analysis
to construct and assess a very large sentiment
Sentiment analysis is one of the most vigorous lexicon including about 120k Arabic terms.
research areas in NLP research field that fo- They put forward an approach that enables them
cuses on analyzing people’s opinions, sentiments, to utilize an available English sentiment lexicon. To
attitudes, and emotions towards several entities evaluate their lexicon, the authors made use of a
such as products, services, organizations, issues, pretreated and labeled dataset of 300 tweets and
events, and topics [34]. A large number of reached an accuracy rate of 87%.
research works on sentiment analysis have been In their attempt to construct a new Arabic lexicon,
recently published in different languages. Hence, [59] proposed a subjectivity and sentiment analysis
to achieve this, we first laid the foundation for system for Egyptian tweets. They built an Arabic
research on the sentiment analysis by reviewing lexicon by merging two modern standard Arabic
relevant literature on past studies conducted in this lexicons (called MPQA and ArabSenti) with two
field. According to this study, sentiment analysis Egyptian Arabic lexicons. The new lexicon is
works fall into three major approaches, namely a composed of 300 positive tweets, 300 negative
machine learning based approach, a knowledge tweets and 300 neutral tweets, for a total of 900
based approach and a hybrid approach. tweets achieving an accuracy of 87%.
The next subsection highlights the related works
of Arabic and Arabic dialects sentiment analysis. 2.3 Hybrid Approach
2.1 Machine Learning Based Approach The Hybrid approach is a combination of the
two approaches already mentioned above. [60]
Machine learning helps data analysts build developed a semantic model called ATSA (Arabic
a model with a large amount of pre-labeled Twitter Sentiment Analysis) based on supervised
words or sentences in order to tackle the machine learning approaches, namely Naı̈ve
classification problem. [54] examined opinion Bayes, support vector machine and semantic
analysis of the Tunisian dialect. Their corpus was analysis. They also created a lexicon by
made up of 17k comments. relying on available resources, such as Arabic
WordNet. The model’s performance has been free, I will not take them] can be converted to
improved compared to the basic bag-of-words Arabizi to become: ”7ata blech manhezhomch
representation with 4.48% for the support vector 5iit”, here the author used the numbers ”7” and ”5”
machine classifier and 5.78% for the classifier which successively replaced the Arabic letters h
NB. In another study conducted by [61], a hybrid and p.
approach combining supervised learning and
rules-based methods was applied for sentiment
intensity prediction. In addition, the authors utilized Moreover, our Arabizi corpus is characterized
not only well-defined linear regression models to by the phenomenon of code switching which
generate scores for the tweets, but also a set of is defined in [36] as “the mixing, by bilinguals
rules from the pre-existing sentiment lexicons to (or multilinguals), of two or more languages in
adjust the resulting scores using Kendall’s score of speech, often without changing the speaker or
about 53%. subject”. This phenomenon is the passage from
one language to another in the same conversation.
For example, ”solde waktech youfa” / [When will the
3 Tunisian Dialect Corpus Description promotion expire?]. This phenomenon appeared
in our Arabizi corpus which includes foreign words
The existence of a corpus is mandatory for a of French origin, such as ”promotion” [promotion],
precise analysis of sentiments, because it is used ”bonjour” [hello], ”caissière” [cashier], etc.
to train and evaluate the models developed. In our
case, we need a corpus in Tunisian dialect for the
One of the major characteristics of Arabizi
opinions analysis in the field of supermarkets. Due
corpus is the presence of abbreviated of foreign
to the lack of available public datasets of this field,
words. Using abbreviated foreign words is one
in this research project, we built our dataset.
of the major characteristics of our Arabizi corpus.
During the data collection phase, two types of Taking the example of the word ”qqles” instead of
corpus are utilized. The comments in the first ”quelques” [a few]. Apart from abbreviations, some
type are written based on a Latin script (also spelling mistakes can be detected from the internet
called Arabizi). The comments in the second users’ performances of some foreign words. These
type of corpus, however, are written by using an mistakes are due to the users’ low French language
Arabic script. As mentioned above, this is due to proficiency. Take the example of the word ”winou el
the habit and ease of writing in Latin, especially cataloug” instead of ”winou el catalogue” [where is
that Tunisians often introduce French words into the catalog?]. Here, we notice that Facebook users
their writings and conversations. Based on these get used to writing words as they listen to them.
characteristics, we decided to divide the corpus
into two different parts, namely Arabizi corpus
and Arabic corpus. This section deals with a
breakdown of the datasets used in our work and 3.2 Arabic Corpus
gives an overview of the corpus statistics.
At this level, our analysis of the data was carried 4.3.3 Light Stemming
out at the aspect level in the Arabic corpus.
Words written in the Tunisian Dialect are often
4.2 Collection of Dataset made up of more than one word; hence the
importance of the root word task. Light rooting
aims at removing all prefixes and suffixes from the
We selected the social network “Facebook” by
word
and keeping its root. For example, the word
concentrating on the supermarket field in Tunisia
in order to gather suitable comments. We
AJ KPA ªÓ mgAztnA / [our supermarket] is changed to
[supermarket] by removing the suffix AK .
èPA ªÓ
have identified five official pages of supermarkets
(Carrefour, Magasin general, Monoprix, Aziza and
Géant). This corpus construction is based on
two important tools: ”Facepager” 1 and ”Export 4.4 Corpus Annotation
Comments” 2 .
Since the primary goal of supervised learning is
to determine the polarity of opinions in advance,
4.3 Processing of the Dataset
we move forward to another crucial step, known
as the annotation of the corpus after pretreating
Once the corpus is collected, the available the two corpora. Manual annotation is therefore
resources must go through a preprocessing stage necessary to build a learning corpus for sentiment
in order to create a usable corpus. analysis. Indeed, our corpus was annotated by
1 https://fr.freedownloadmanager.org/Windows- native Tunisian Dialect speakers who were asked
PC/Facepager-GRATUIT.html to classify the comments according to already
2 https://exportcomments.com/ well-defined categories and sentiment classes.
4.4.1 Annotation for Sentiment Analysis at the Ðñ @ Y ®K . [how much?] , AJËñ kP [Please
Aspect Level
ñÓQK . [promotion], etc.
reduce the price] , àñJ
the annotation of the aspect-based sentiment
— Quality: This refers to an advice via a
analysis depends only on the Arabic script corpus,
product or a service. For example: J
ªK [it’s
in which we first labeled our collected comments
with five distinct classes. The first class is ”very disgusting], J® K [that’s wonderful] , etc.
positive”. It’s used when the comment contains
words that express total satisfaction with a service — General: The user does not express his
or a product. For example, J®K [very nice]. The opinion clearly. It is difficult to understand
second class ”positive” is used if the comment whether his opinion is related to either the
expresses a positive feeling, such as satisfaction, price of the product or the quality of the
enthusiasm, etc. For example, éJ
K. [delicious]. product.
you work during Eid] is informative with no As regards the method of analyzing opinions at the
word of sentiment. Negative means if the sentence level, the two corpus have been labelled
comment expresses a negative feeling, such into five classes which are shown in table 3.
as dissatisfaction, regret or any other negative
feelings. For example, á
J.K
Ag Qå
AK
[they are not
4.5 Feature Vectors
beautiful], etc. Finally, very negative means when
the comment expresses a very negative feeling, Our method applied two main feature vectors,
such as annoyance, disappointment or any other namely word embedding vectors and the morpho-
very negative feelings. For example, the word syntactic analysis. Although the preprocessing
J J . j.«AÓ [very bad]. step was performed in the two corpora, the
Then, we annotated our Arabic corpus with comments are not ready to be used by two
20 already well-defined categories. Each neural network algorithms because they require a
category comprises two major parts that form representation of the words that were considered
the tuple ”E#A”, namely the aspect entity (E) as feature vectors. Our classifiers, CNN, LSTM
and the attribute of aspect entity (A). The list and Bi-LSTM deep learning networks, take as input
of aspect entities is composed of 9 aspects, the vector representations of each word.
namely, “drinks”, “aliments”, “services”,” locations”, Indeed, it is necessary to go through the creation
“cleaners”, “electronics”, “utensils” and “others”. of a characteristic vector for each word in a
However, the list of attributes of aspect entities comment. Consequently, we implemented the
is, namely, “general”, “quality and “price”. These Gensim 3 of the Word2Vec model. This model
attribute are defined for each entity, except for the plays a key role as it uses a deep neural network,
two entities; “locations” and “services”. Thus, we manipulates sentences in a given document,
fixed a single entity attribute “general” for these last and generates output vectors. The Word2Vec
two entities. Table 2 reports aspect categories with model was developed by Google researchers led
examples of topics discussed in each category. by [37]. The word2vec algorithms include two
different models: skip-gram and the Continuous
Aspect entity attributes corresponding to entity
Bag-of-Words (CBOW) model. In the first model,
labels are shown as follows:
the distributed representations of the input word
are used to predict the context. In the second
— Price: this attribute includes promotions and
payment facilities. For example, the words 3 https://radimrehurek.com/gensim/models/word2vec.html
Table 3. Sentiment classes with examples of comments from the two corpus
les caissiere hala yahkyou m3ana
bkelet tourbya
very pos-
é«ðP ð@ð a7ssen afar wahsen kadya w
itive service fi mg
very neg- Aî DÓ
Qå AÓ éÓñm
Ì '@ PA¢«
áÓ úΫ@ ©J
.K úÍ@ èQK
Q«
7asilou allah latraba7kom ou lat-
ative
far7kom
Neutral ñÓAJË@ úΫ ú
æ ® K ú
Í@ éJJ
»AÜ Ï @ ù
ë ø
Yë famech rakadha on promotion svp
model, however,the representations of the context namely, sentence-based sentiment analysis and
are combined to predict the word in the middle. aspect-based sentiment analysis.
For the morpho-syntactic analysis, we employed
We applied the Word2vec algorithm as input for
the tool developed by [52], who proposed a method
both neural networks. The general architecture of
to remove the ambiguity in the output of the
neural networks is illustrated in figure 2.
morphological analyzer from the Tunisian Dialect.
They disambiguated results of the Al-Khalil-TUN The neural network is defined by determining the
Tunisian Dialect morphological analyzer [53] Their number of input layers, the number of hidden layers
suggested method achieved an accuracy of 87%. and the number of output layers. As illustrated in
For our work, entity names are identified by words Figure 2, the neural network architecture can be
which are labeled with ”noun” and ”prop-noun”. expressed in this way:
about sentiment words, they are recognized by
words labeled with the words ”adj” and ”verb”. In the input layer, the inputs (1 · · · n) indicate a
sequence of words. Then, i = (i1 · · · in ) present
the vector representations of entries entered. The
4.6 Classification
neural network consists of h = (h1 · · · hn ) hidden
For the implementation of these methods, we layers that aim to map the vectors i to hidden layers
selected three algorithms, namely CNN, LSTM h. The last layer, the output layer, receiving the last
and Bi-LSTM by relaying the deep learning hidden layer output to merge and produce the final
methods in order to implement the two tasks, result o.
4.6.1 Convolution Neural Network model (CNN) dimensional representation of input data generated
by the convolution layer. The output from the
CNN is also called Convnets or Cnns, it is one of pooling layer is passed to a fully connected softmax
the models of deep learning that marks impressive layer. As a multi-class classification was adopted
results in the field of TALN in general, and in the in our work, we used a Softmax classification layer
analysis of feelings. which predicts the probabilities for each class.
Our CNN model consists of an input and an The figure 3 summarizes how the CNN model
output layer, along with numerous hidden layers. works for sentiment analysis at the two levels.For
Conventionally, the layers of a neural network are the aspect level, for the aspect category model, the
fully connected. This explains the fact that the input is a vector representation of each sentence
output of each layer is the input of the next layer. with its aspect words, while, for the sentiment
The hidden layers consist of convolution layers, model, is a vector representation of each sentence
pooling layers, two fully connected layers, and with its sentiment words. Concerning the output,
activation functions. The pooling layer is applied for the aspect category model, the output is a list
to the output of the convolution layer. The fully of 20 categories classes, and, for the sentiment
connected layer aims to concatenate all the vectors model, it’s a list of 5 classes. While, for the
into one single vector. The activation function sentence level, the input of the model is a vector
introduces non-linearity into the neural network representation of each sentence. The output is a
through a softmax classification layer. The output list of 5 classes.
layer generates the polarity of each input comment.
To summarize, to train our CNN model, 4.6.2 Long Short-Term Memory Model (LSTM)
we adopted the Word2vec algorithm for the
representation of words with a size of 300. The RNN Model is characterized by the sequential
This representation presents the input of the aspect of an entry in which the word order is of a
convolution layer. The convolution results are significant importance. In addition, the RNN gives
grouped or aggregated to a representative number the possibility of processing variable length entries.
via the pooling layer. This number is sent to In this work, long short-term memory (LSTM) is
a fully connected neural structure in which the a particular type of neural network that is used to
classification decision is based on the weights learn sequence data. As a type of RNN, LSTM
assigned to each entity in the text. reads the input sequence from left to right. It is
Indeed, the main purpose of the fully connected also capable of learning long-term relationships.
layer is the reduction or compression of the That is why the prediction of an input depends on
the anteposed or postposed context. So, RNN is backward layer by receiving both status information
designed to capture long distance dependencies. from the previous sequence (backward) and the
The entry into the LSTM model is a sequence next sequence (forward). Bi-LSTM is therefore
of word representations using the Word2vec very useful where the context of the entry is
algorithm. Then, those representations are necessary, for example when the word negation
passed to an LSTM layer. The output of this appears before a positive term.
layer is also passed to a softmax activation
layer which produces predictions on the whole In our work, we used a one-layer bi-LSTM where
vocabulary words. the entry of the model is a sequence of words M
represented using the Word2vec algorithm. Then,
The figure 4 summarizes how the LSTM model
M is passed to a Bi-LSTM layer, the output of this
works for sentiment analysis at the two levels.
layer is passed to a softmax activation layer which
As mentioned above in the CNN model, For the
produces predictions on the whole vocabulary, for
aspect level, the input is a vector representation of
each step of the sequence. Each method of CNN
each sentence with its aspect words and sentiment
and RNNs in general has its characteristics and
words. The output of this level is a list of 20
differs from the other.
categories classes and other 5 sentiment classes.
For the sentence level, the input of the model is a RNNs process entries sequentially, while CNNs
vector representation of each sentence. While, the process information in parallel. In addition, CNN
output is a list of 5 classes. models are limited to a fixed length entry, while
RNN models have no such limitation. The major
4.6.3 Bidirectional Long Short-Term Memory disadvantage of RNNs is that they train slowly
Model (Bi-LSTM) compared to CNNs.
The LSTM network reads the input sequence from The architecture of the Bi-LSTM model for the
left to right, while Bi-LSTM which is variant of two levels, sentence and aspect level is shown in
LSTM, it relies on the connection of two LSTMs Figure 5. As mentioned above in the CNN and
layers of reverse directions (forward and backward) LSTM models, the input of the aspect level is a
on the input sequence. The first layer is the vector representation of each sentence with its
input sequence and the second layers is the aspect words and sentiment words. The output of
reversed copy of the input sequence The output this level is a list of 20 categories classes and other
layer combines the outputs of the forward layer and 5 sentiment classes.
For the sentence level, the input of the model is a we made use of Al-Khalil-TUN Tunisian Dialect
vector representation of each sentence. The output morphological analyzer developed by [52] to
generating is a list of 5 classes. remove the output ambiguity.
Table 4. The corpus size followed by the vocabulary size for sentence-based sentiment analysis
20 aspect category classes of our Arabic corpus precision and recall measures. The F-Measure
according to the estimated number of occurrences. also called F-score is computed as the harmonic
The axis refers to the aspect categories in mean of the precision and recall with values
our dataset. x shows the estimated number of ranging between 0 (representing the worst score)
occurrences. The number of comments classified and 1 (representing the best score). Therefore, the
as (others-general) is the largest part of the F-Measure is calculated as follows:
dataset.
2 × (P recision × Recall)
The figures 7 and 8 presents some statistics F − M easure = . (1)
(P recision + Recall)
of sentiment classes for both, Arabic corpus and
Arabizi corpus.
5.3 Evaluation Results and Discussion
5.2 Evaluation Metrics
In this part we will present the outcomes
In order to assess the performance of the two of the evaluations carried out for the deep
levels, at sentence level and aspect level. Con- learning models (CNN, LSTM and Bi-LSTM). The
cerning the aspect level,for the aspect category performances of those models are represented
model and the sentiment model, we calculated the by using F-Measure measurement. At aspect
F-Measure 1 of each model by combining both level, we present the F-Measure of each model
Table 5. The corpus size followed by the vocabulary size for aspect-based sentiment analysis
depending on the stop words and stemming used classification, we made some modifications. In
in the pretreatment phase. this vein, the “very positive” class is replaced by
Table 6 shows F-Measure values of both the “positive” class, and the “very negative” class
the aspect category model and the sentiment is changed to the “negative” class. Finally, for a
model. According to table 6, we noticed that binary classification, we went through the same
the use of stop word removal and stemming had steps of the three-way classification by eliminating
almost no improvements for the three models. the neutral class.
For this reason, we decided to carry out the According to table 7, the best F-Measure
other experiments by ignoring stemming and reached 49% with 20 categories and 5 sen-
stop-word removal. timent classes for the aspect category model
To achieve better results, we will test the using Bi-LSTM model. However, the highest
performance of our models by classifying our F-measure was 78% for the sentiment model using
experiences into 8 distinct groups. We will try Bi-LSTM model.
to modify the number of aspect categories and In the table 8, we will focus on the modified
sentiment classes. To do this, we changed the number of aspect categories. We will also evaluate
number of sentiment classes (5, 4, 3 and 2) our models based on each sentiment class (5, 4, 3
without modifying the number of aspect categories and 2). Indeed, we modified the aspect categories
presented in table 7. For a four-way classification, by removing the list of attributes of aspect entities
we ignored the neutral class. Fora three-way (“general”, “quality and “price”).
Table 7. F-Measures of the aspect category model and sentiment model with 20 aspect categories and different
sentiment classes
Table 8. F-Measure of aspect category model and sentiment model with different sentiment classes and 8 aspect
categories
Table 9. F-Measure values for the three deep learning railway stations. Our constructed corpus, however,
models was gathered from the official supermarket pages.
Hence, all these causes have a negative impact on
CNN LSTM Bi-LSTM
the performance of our developed classifiers.
With 5 classes 66% 68% 69%
At sentence level, table 8 shows F-Measure
With 4 classes 71% 72% 73%
values for the three deep learning models.
With 3 classes 69% 87% 72%
In order to achieve better results, we will test
With 2 classes 86% 87% 87%
the performance of our models by classifying our
experiences into 4 distinct groups by changed
Based on the outcomes displayed in table 8, the number of sentiment classes (5, 4, 3 and
the CNN-based classification and Bi-LSTM-based 2). For a four-way classification, we ignored the
classification achieved best F-Measures values neutral class. For a three-way classification, we
with 62% for the aspect category model. For made some modifications. In this vein, the “very
the sentiment model, the best F-Measure value positive” class is replaced by the “positive” class,
obtained was 78% using CNN model. and the “very negative” class is changed to the
“negative” class. Finally, for a binary classification,
The result obtained by the two models was we went through the same steps of the three-way
not good for several reasons. First, deep classification by eliminating the neutral class.
learning methods need a large dataset for best According to table 9, the highest F-Measure
performance. However, this is not the case for reached was 87% with LSTM and Bi-LSTM with
our work as our corpus size is only 17k. Second, three-way and two-way classification using 5
morpho-syntactic analyzer that produces many sentiment classes.
empty lines was not able to extract all sentiment
words and aspects.
5.4 Discussion
Moreover, the training corpus utilized to train
the morpho-syntactic analyzer [52] was a spoken In this section, for the sentence level, we will
a corpus that consists of not only radio and TV compare the results obtained by our classifiers with
broadcasts but also conversations recorded in the work [69] and of [70].
Table 10. Comparison of CNN, LSTM and Bi-LSTM results for sentiment analysis at sentence level
absence of a standard orthography, people use a 8. Darwish, K. (2013). Arabizi detection and conver-
spontaneous orthography based on phonological sion to Arabic. ANLP-EMNLP.
criteria. Fourth, the Tunisian Dialect is usually 9. Moussa, M.E., Mohamed, E.H., Haggag, M.H.
written without diacritical signs. One of the (2018). A survey on opinion summarization tech-
major functions of these signs is to determine niques for social media. Future Computing and
and facilitate the meaning of words, phrases Informatics Journal.
or sentences.
10. Hu Ya-Han, Chen Yen-Liang, Chou Hui-Ling
Our future works will focus, first, on improve and (2017). Opinion mining from online hotel re-
develop our models to be more precise in detecting views: A text summarization approach. Information
negation and to deal with both the sarcasm Processing Management, Vol. 53, pp. 436–449.
problem and spam detection. In addition, we will DOI:10.1016/j.ipm.2016.12.002.
try to increase the dataset size by transliterating
11. Qwaider, C., Saad, M., Chatzikyriakidis, S.,
the Arabizi corpus into Arabic. Dobnik, S. (2018). Shami: A corpus of Levantine
Finally, in order to ameliorate the outcomes Arabic dialects. Proceedings of the Eleventh
of our models (CNN, LSTM and Bi-LSTM), we International Conference on Language Resources
will try to test other deep learning techniques, and Evaluation (LREC-2018).
like DNN (deep neural networks) and DBN (deep 12. Jiménez-Zafra, S.M., Martı́n-Valdivia, M.T.,
belief networks). Martı́nez-Cámara, E., Ureña-López, L.A. (2016).
Combining resources to improve unsupervised
sentiment analysis at aspect-level.
References
13. Noferesti, S., Shamsfard, M. (2015). Resource
1. Liu, B., Zhang, L. (2012). A Survey of Opinion construction and evaluation for indirect opinion
Mining and Sentiment Analysis. Mining Text Data, mining of drug reviews. PloS One, Vol. 10, No. 5.
pp. 415–463. DOI: 10.1007/978-1-4614-3223-4 13. DOI: 10.1371/journal.pone.0124993.
2. Liu, B. (2006). Web data mining: exploring 14. Dua, M., Nanda, Ch., Nanda, G. (2018). Sentiment
hyperlinks, contents, and usage data. Springer. analysis of movie reviews in Hindi language using
machine learning. Conference: International Con-
3. Priyadarshana, Y.H.P.P., Gunathunga, ference on Communication and Signal Processing
K.I.H., Perera, K.K.A.N.N., Ranathunga, L., (ICCSP), India. DOI:10.1109/ICCSP.2018.8524223.
Karunaratne, P.M., Thanthriwatta, T.M. (2015).
Sentiment analysis: Measuring sentiment strength 15. Varathan, K.D., Giachanou, A., Crestani, F.
of call centre conversations. Proceeding IEEE (2015). Comparative opinion mining: A review. J.
ICECCT, pp. 1–9. Assoc. for Inf. Sci. Technol, Vol. 68, No. 4,
pp. 811–829.
4. Nitin, J., Liu, B. (2006). Mining comparative
sentences and relations. Proceedings of National 16. Nabil, M., Aly, M., Atiya, A. (2015). ASTD:
Conf. on Artificial Intelligence (AAAI-2006). Arabic sentiment tweets dataset. Proceedings of
5. Noferesti, S., Shamsfard, M. (2015). Resource the Conference on Empirical Methods in Natural
construction and evaluation for indirect opinion Language Processing, pp. 2515-–2519.
mining of drug reviews. PloS One, Vol. 10, No. 5, 17. Van de Kauter, M., Breesch, D., Hoste,
e0124993. V. (2015). Fine-grained analysis of explicit and
6. Cambria, E., Poria, S., Bisio, F., Bajpai, R., implicit sentiment in financial news articles. Expert
Chaturvedi, I. (2015). The CLSA model: A novel Systems with Applications, Vol. 42, No. 11.
framework for concept-level sentiment analysis. DOI:10.1016/j.eswa.2015.02.007.
LNCS, Vol. 9042, pp. 3–22. Springer.
18. Tang, D., Qin, B., Liu, T. (2015). Document
7. Manna, S. (2018). Sentiment analysis based on modeling with gated recurrent neural network
different machine learning algorithms. International for sentiment classification. Proceedings of the
Journal of Computer Sciences and Engineering, Conference on Empirical Methods in Natural
Vol. 6, No. 6, pp. 1116–1120. Language Processing, pp. 1422–1432.
19. Mankar, S.A., Ingle, M. (2015). Implicit sentiment 29. Fsih, E., Boujelbane, R., Belguith, L. (2018).
identification using aspect based opinion mining. Tunisian dialect resources for opinion analy-
DOI:10.17762/ijritcc2321-8169.150491. sis on social media. JCCO Joint International
Conference on ICT in Education and Training,
20. Wiebe, J., Wilson, T., Bruce, R.F., Bell,
International Conference on Computing in Arabic,
M., Martin, M. (2004). Learning Subjective
and International Conference on Geocomputing
Language. Computational Linguistics, Vol. 30, pp.
(JCCO: TICET-ICCA-GECO). DOI:10.1109/ICCA-
277–308.
TICET.2018.8726218.
21. Wiebe, J., Riloff, E. (2005). Creating subjective
30. Chaturvedi, I., Cambria, E., Welsch, R.,
and objective sentence classifiers from unannotated
Herrera, F. (2018). Distinguishing between facts
texts. Computational Linguistics and Intelligent Text
and opinions for sentiment analysis: Survey and
Processing, 6th International Conference, CICLing,
challenges. Information Fusion, Vol. 44, pp. 65–77.
pp. 486. DOI: 10.1007/978-3-540-30586-6 53.
DOI:10.1016/j.inffus.2017.12.006.
22. Shubham, D., Mithil, P., Shobharani, M.,
31. Alharabi, A., Taileb, M., Kalkatawi, M.
Subramanain, S. (2017). Aspect level sentiment
(2019). Deep learning in Arabic sentiment
analysis using machine learning. IOP Conference
analysis: An overview, journal of information
Series Materials Science and Engineering, Vol. 263,
science. Journal of Information Science.
No. 4. DOI: 10.1088/1757-899X/263/4/042009.
DOI:10.1177/0165551519865488.
23. Oueslati, O., Cambria, E., HajHmida, M.B.,
32. Refaee, E., Rieser, V. (2014). An Arabic twitter
Ounelli, H. (2020). A review of sentiment analysis
corpus for subjectivity and sentiment analysis.
research in Arabic language. Future Generation
LREC, pp. 2268–2273.
Computer Systems, Vol. 112, pp. 408–430. DOI:
10.1016/j.future.2020.05.034. 33. Oueslati, O., Ben Hajhmida, M., Ounelli,
H., Cambria, E. (2019). Sentiment analysis of
24. Kim, S.M., Pantel, P., Chklovski, T., Pen- influential messages for political election forecast-
nacchiotti, M. (2006). Automatically assessing ing. International Conference on Computational
review helpfulness. EMNLP’06: Proceedings of Linguistics and Intelligent Text Processing.
the Conference on Empirical Methods in Natural
Language Processing, pp. 423–430. Association for 34. Oueslati O., Ahmed, I.S.K., Habib O. (2018).
Computational Linguistics. Sentiment Analysis for Helpful Reviews Prediction.
International Journal of Advanced Trends in
25. Bisio, F., Gastaldo, P., Peretti, Ch., Zunino, R., Computer Science and Engineering, Vol. 7, No. 3,
Cambria, E. (2013). Data intensive review mining pp. 34–40. DOI:10.30534/ijatcse/2018/02732018.
for sentiment classification across heterogeneous
domains. Advances in Social Networks Analysis 35. Liu, B., Zhang, L. (2012). A Survey of Opinion
and Mining (ASONAM), pp. 1061–1067. DOI: Mining and Sentiment Analysis. Mining Text Data,
10.1145/2492517.2500280. pp. 415–463. DOI:10.1007/978-1-4614-3223-4-13.
26. Ofek, N., Poria, S., Rokach, L., Cambria, E., 36. Poplack, S. (2001). Code-switching (linguistic).
Hussain, A., Shabtai, A. (2016). Unsupervised 37. Mikolov, T., Chen, K., Corrado, G., Dean, J.
commonsense knowledge enrichment for domain- (2013). Efficient estimation of word representations
specific sentiment analysis. Cognitive Computation, in vector space. arXiv:1301.3781.
Vol. 8, No. 3, pp. 467–477.
38. Masmoudi, A., Habash, N., Khmekhem, M.,
27. Hammad, A.S.A., El-Halees, A. (2013). An ap- Estéve, Y., Belguith, L. (2015). Arabic translit-
proach for detecting spam in Arabic opinion reviews. eration of romanized Tunisian dialect text: A
The International Arab Journal of Information preliminary investigation. Computational Linguis-
Technology, Vol. 12. tics and Intelligent Text Processing, 16th In-
28. Ghose, A., Panagiotis I. (2007). Designing novel ternational Conference, CICLing, pp. 608–619.
review ranking systems: predicting the usefulness DOI:10.1007/978-3-319-18111-0 46.
and impact of reviews. Proceedings of the Ninth 39. Abir Masmoudi (2016). Approche hybride pour la
International Conference on Electronic Commerce reconnaissance automatique de la parole pour la
(ICEC ’07), Association for Computing Machinery, langue arabe. Thèse de doctorat en informatique,
pp. 303–310. DOI:10.1145/1282100.1282158 Université de sfax.
40. Masmoudi, A., Mdhaffar, S., Sellami, R., of Computer Science and Information Technology,
Belguith, L. (2019). Automatic diacritics restoration IJCSMC, Vol. 4, No. 9, pp. 212–219.
for Tunisian dialect. ACM Transactions on Asian and
50. Xing, F., Pallucchini, F., Cambria, E. (2019).
Low-Resource Language Information Processing,
Cognitive-inspired domain adaptation of
Vol. 18, No. 3, pp. 1–18. DOI:10.1145/3297278.
sentiment lexicons. Information Processing
41. Masmoudi, A., Khmekhem, M.E., Khrouf, M., and Management, Vol. 56, No. 3, pp. 554–564.
Belguith, L.H. (2020). Transliteration of Arabizi into DOI:10.1016/j.ipm.2018.11.002.
Arabic script for Tunisian dialect. ACM Transactions
on Asian and Low-Resource Language Information 51. Zribi, I., Boujelbane, R., Masmoudi, A., El-
Processing, Vol. 19, No. 32. DOI:10.1145/3364319. louze, M., Belguith, L., Habash, N. (2014).
A Conventional Orthography for Tunisian Arabic.
42. Xing, F., Pallucchini, F., Cambria, E. (2019).
Proceedings of the Ninth International Conference
Cognitive-inspired domain adaptation of
on Language Resources and Evaluation: LREC’14,
sentiment lexicons. Information Processing
pp. 2355–2361.
and Management, Vol. 56, No. 3, pp. 554–564.
DOI:10.1016/j.ipm.2018.11.002. 52. Zribi, I., Ellouze, M., Belguith, L., Blache, Ph.,
43. Younes, J., Achour, H., Souissi, E., Ferchichi, (2017). Morphological disambiguation of Tunisian
A. (2018). Survey on corpora availability for the dialect. Journal of King Saud University-Computer
Tunisian dialect automatic processing. JCCO Joint and Information Sciences, Vol. 29, No. 2, pp.
International Conference on ICT in Education 147–155. DOI:10.1016/j.jksuci.2017.01.004.
and Training, International Conference on Com- 53. Zribi, I., Ellouze, M., Belguith, L. (2013). Mor-
puting in Arabic, and International Conference phological Analysis of Tunisian Dialect. International
on Geocomputing (JCCO: TICET-ICCA-GECO). Joint Conference on Natural Language Processing,
DOI:10.1109/icca-ticet.2018.8726213. pp. 992–996.
44. Pang, L. et Vaithyanathan (2002). Thumbs
54. Medhaffar, S., Bougares, F., Estève, Y.,
up? Sentiment Classification using Machine
Hadrich-Belguith, L. (2017). Sentiment analysis
Learning Techniques. Proceedings of the
of Tunisian dialects: Linguistic resources and
Conference on Empirical Methods in Natural
experiments. Proceedings of the Third Arabic
Language Processing (EMNLP), pp. 79–86.
Natural Language Processing Workshop, pp.
DOI:10.3115/1118693.1118704.
55–61. DOI:10.18653/v1/W17-1307.
45. Boukadida, N. (2008). Connaissances
phonologiques et morphologiques dvationnelles 55. Al-Ayyoub, M., Bani-Essa, S., Alsmadi, I. (2015).
et apprentissage de la lecture en arabe (Etude Lexicon-based sentiment analysis of Arabic tweets.
longitudinale). Thèse de doctorat, Université Int. J. Social Network Mining, Vol. 2, No. 2, pp. 101–
Rennes. 114. DOI:10.1504/IJSNM.2015.072280.
46. Peng, H., Ma, Y., Li, Y., Cambria, E. (2018). 56. Gupta, A., Pruthi, J., Sahu, N. (2017). Sentiment
Learning multi-grained aspect target sequence analysis of tweets using machine learning ap-
for chinese sentiment analysis. Knowledge-Based proach. International Journal of Computer Science
Systems, Vol. 148, pp. 167–176. and Mobile Computing, IJCSMC, Vol. 6, No. 4, pp.
444–458.
47. Wahsheh, H., Al-Kabi, M., Alsmadi, I. (2013).
Spar: A system to detect spam in Arabic opinions. 57. Kwaik, K.A., Saad, M., Chatzikyriakidis, S.,
IEEE Jordan Conference on Applied Electrical Engi- Dobnik, S. (2019). LSTM-CNN Deep learning
neering and Computing Technologies (AEECT), pp. model for sentiment analysis of dialectal Arabic.
1–6. DOI:10.1109/AEECT.2013.6716442. Arabic Language Processing: From Theory to
48. Gautami T., Naganna, S. (2015). Feature se- Practice, pp. 108–121.
lection and classification approach for senti- 58. Mohamed Amine Jerbi, Hadhemi Achour, Emna
ment analysis. Machine Learning and Applica- Souissi (2019). Sentiment analysis of code-
tions: An International Journal (MLAIJ). DOI: switched Tunisian dialect: Exploring RNN-based
10.5121/mlaij.2015.2201. techniques. Arabic Language Processing: From
49. Preety, Dahiya, S. (2015). Sentiment analysis Theory to Practice, pp. 122–131. DOI:10.1007/978-
using SVM and naive Bayes algorithm. Journal 3-030-32959-4 9.
59. El-Makky, N., Nagi, K., El-Ebshihy, A., Apady, 69. Heikal Maha, Torki Marwan, El-Makky
E., Hafez, O., Mostafa, S., Ibrahim, S. (2015). Nagwa (2018). Sentiment analysis of
Sentiment Analysis of Colloquial Arabic Tweets. Arabic tweets using deep learning. Procedia
The 3rd ASE International Conference on Social Computer Science, Vol. 142, pp. 114–122.
Informatics (SocialInformatics). DOI:10.1016/j.procs.2018.10.466.
60. Alowaidi, S., Saleh, M., Abulnaja, O.A. (2017). 70. Al-Azani, S., El-Alfy, E.S.M. (2017). Hybrid deep
Semantic sentiment analysis of Arabic texts. learning for sentiment polarity determination of
International Journal of Advanced Computer Arabic microblogs. Liu, D., Xie, S., Li, Y., Zhao, D.,
Science and Applications, Vol. 8, No. 2. El-Alfy, E.S. (eds) Neural Information Processing
DOI:10.14569/IJACSA.2017.080234. ICONIP. Lecture Notes in Computer Science, Vol.
61. Refaee, E., Rieser, V. (2016). iLab-Edinburgh at 10635, pp. 491–500.
SemEval-2016 Task 7: A Hybrid Approach for
71. Hossam, S.I., Mervat, G., Sherif, M.A. (2015).
Determining Sentiment Intensity of Arabic Twitter
Sentiment Analysis for Modern Standard Arabic
Phrases. Proceedings of the 10th International
and Colloquial. International Journal on Natural
Workshop on Semantic Evaluation (SemEval), pp.
Language Computing IJNLC, Vol. 4, No. 2.
474–480. DOI:10.18653/v1/S16-1077.
62. Younes, J., Achour, H., Souissi, E. (2015). 72. Barhoumi, A., Estève, Y., Aloulou, Ch., Belguith,
Constructing linguistic resources for the Tunisian L.H. (2017). Document Embeddings for Arabic
dialect using textual user-generated contents on Sentiment Analysis. LPKM.
the social web. International Conference on Web 73. Al-Smadi, M., Al-Ayyoub, M., Jararweh,
Engineering, pp. 3–14. Y., Qawasmeh, O. (2019). Enhancing
63. McNeil, K. (2011). Tunisian Arabic corpus: Creating aspect-based sentiment analysis of Arabic
a written corpus of an unwritten language. hotels’ reviews using morphological, syntactic and
International Symposium on Tunisian and Libyan semantic features. Information Processing and
Arabic Dialects, University of Vienna. Management, Vol. 56, No. 2, pp. 308–319. DOI:
10.1016/j.ipm.2018.01.006.
64. Habash, N.Y. (2010). Introduction to
Arabic natural language processing. 74. Wagh, B., Shinde, Kale, P.A. (2018). A twitter
Synthesis Lectures on Human Language sentiment analysis using NLTK and machine learn-
Technologies, Vol. 3, No. 1, pp. 1–187. ing techniques. International Journal of Emerging
DOI:10.2200/S00277ED1V01Y201008HLT010. Research in Management and Technology, Vol. 6,
65. Hamdi, A., Gala, N., Nasr, A. (2014). Automatically No. 37. DOI: 10.23956/ijermt.v6i12.32.
building a Tunisian lexicon for deverbal nouns. 75. Al-Shalabi, R., Kanan, G.G., Gharaibeh, M.H.
Proceedings of the First Workshop on Applying NLP (2006). Arabic text categorization using kNN
Tools to Similar Languages, Varieties and Dialects, algorithm.
pp. 95–102. DOI:10.3115/v1/W14-5311.
76. Masmoudi, A., Khmekhem, M.E., Estève, Y.,
66. Graja, M., Jaoua, M., Belguith, L.H. (2010).
Belguith, L.H., Habash, N. (2014). A corpus
Lexical study of a spoken dialogue corpus in
and phonetic dictionary for Tunisian Arabic speech
Tunisian dialect. Proceedings of the Interna-
recognition. Proceedings of the Ninth International
tional Arab Conference on Information Technology,
Conference on Language Resources and Evalua-
Benghazi-Libya.
tion (LREC’14), pp. 306–310.
67. Abir Masmoudi (2016). Approche hybride pour la
reconnaissance automatique de la parole pour la 77. Masmoudi, A., Estève, Y., Khmekhem, M.E.,
langue arabe. Thèse de doctorat en informatique, Bougares, F., Belguith, L.H. (2014). Phonetic tool
Université de sfax. for the Tunisian Arabic. SLTU, pp. 253–256.
68. Al-Azani, S., El-Alfy, E. (2018). Emojis-based 78. Zribi, I., Boujelbane, R., Masmoudi, A., Ellouze,
sentiment classification of Arabic microblogs M., Belguith, L.H., Habash, N. (2014). A
using deep recurrent neural networks. conventional orthography for Tunisian Arabic.
International Conference on Computing Proceedings of the Ninth International Conference
Sciences and Engineering (ICCSE), pp. 1–6, on Language Resources and Evaluation (LREC’14),
DOI:10.1109/ICCSE1.2018.8374211. pp. 2355–2361.
79. Elsahar, H., El-Beltagy, S. (2015). Building large Technology and Computational Linguistics JLCL,
Arabic multi-domain resources for sentiment analy- Vol. 30 No. 1, pp. 1–25.
sis. Lecture Notes in Computer Science, Vol. 9042,
83. Huang, H.H., Wang, J.J., Chen, H.H. (2017).
pp. 23–34. DOI:10.1007/978-3-319-18117-2 2.
Implicit opinion analysis: Extraction and po-
80. Duwairi, R.M. (2015). Sentiment analysis for larity labelling. Journal of the Association for
dialectical Arabic. Proceedings 6th ICICS Information Science and Technology, Vol. 68.
International Conference on Information DOI:10.1002/asi.23835.
and Communication Systems, pp. 166–170.
DOI:10.1109/IACS.2015.7103221. 84. Al-Twairesh, N., Al-Khalifa, H., Al-Salman,
A. (2014). Subjectivity and sentiment analysis
81. Al-Obaidi, A.Y., Samawi, V. (2016). Opinion of Arabic: Trends and challenges. Computer
mining: Analysis of comments written in Arabic Systems and Applications AICCSA, IEEE/ACS
colloquial. Proceeding of the Conference of World 11th International Conference on, pp. 148–155.
Congress on Engineering and Computer Science DOI:10.1109/AICCSA.2014.7073192.
WCECS, Vol. 1, pp. 470–475.
82. Bayoudhi, A., Ghorbel, H., Koubaa, H., Hadrich-
Belguith, L. (2105). Sentiment classification
at discourse segment level: Experiments on Article received on 07/09/2020; accepted on 29/09/2020.
multi-domain Arabic corpus. Journal for Language Corresponding author is Abir Masmoudi.