A Framework for Arabic Concept-Level Sentiment Analysis using SenticNet

International Journal of Electrical and Computer Engineering (IJECE)
Vol. 8, No. 5, October 2018, pp. 4015~4022
ISSN: 2088-8708, DOI: 10.11591/ijece.v8i5.pp4015-4022  4015
Journal homepage: http://iaescore.com/journals/index.php/IJECE
A Framework for Arabic Concept-Level Sentiment Analysis
using SenticNet
Hend G. Hassan, Hitham M. Abo Bakr, Ibrahim E. Ziedan
Department of Computer and systems Engineering, Zagazig University, Egypt
Article Info ABSTRACT
Article history:
Received Dec 29, 2017
Revised Mar 26, 2018
Accepted Jul 11, 2018
Arabic Sentiment analysis research field has been progressing in a slow pace
compared to English and other languages. In addition to that most of the
contributions are based on using supervised machine learning algorithms
while comparing the performance of different classifiers with different
selected stylistic and syntactic features. In this paper, we presented a novel
framework for using the Concept-level sentiment analysis approach which
classifies text based on their semantics rather than syntactic features.
Moreover, we provided a lexicon dataset of around 69 k unique concepts that
covers multi-domain reviews collected from the internet. We also tested the
lexicon on a test sample from the dataset it was collected from and obtained
an accuracy of 70%. The lexicon has been made publicly available for
scientific purposes.
Keyword:
Arabic reviews
Opinion mining
Sentiment analysis
Sentiment lexicons
Copyright © 2018 Institute of Advanced Engineering and Science.
All rights reserved.
Corresponding Author:
Hend G. Hassan,
Department of Computer and systems Engineering,
Zagazig University, Egypt.
Email: hend.hgh25@gmail.com
1. INTRODUCTION
Sentiment analysis (SA), a form of text classification, is the process of classifying a given
document/paragraph/sentence into two or more classes. SA involves 8 key tasks [1] including subjectivity
detection, where data is classified into subjective or objective data, and polarity detection where subjective
data is further classified into positive, negative or mixed. Those tasks are concerned with internet users‘
public opinions; the data they share help gaining perspective of the overall sentiment about a specific
product, service, person, etc. not only to institutions or companies but also to other internet users. However,
that raw opinionated big data is unstructured and requires semantics and syntactic analysis in order to be
machine understandable.
Existing sentiment analysis approaches are categorized into four main ones [2]: keyword spotting,
lexical affinity, statistical methods and concept-level sentiment analysis. Keyword spotting is basically
spotting keywords in the sentence and classifying it afterward. Keywords, which have positive, negative or
neutral polarity, are clear sentimental words on their own like ‗care‘, ‗angry‘, ‗glad‘, ‗sick‘, etc. However,
using them in a sentence may have a different sentiment other than the one they have on their own. For
example ‗‗I care for the wrong people‘‘, care has positive polarity, but the whole sentence evokes a negative
sentiment causing a misclassification error. Asides to that, the sentence may not include any keywords like
‗‗I would never buy this book‘‘ which indicates a negative opinion about the book yet can't be classified. Or
has a misleading comparison like ‗‗this book is as good as a hole in the head‘‘. Briefly, this approach is
known to be the most naïve one and the most popular too for its ease of implementation and accessibility.
Lexical affinity dives a little deeper in the keywords semantics than the first approach. It assigns
arbitrary words with a probabilistic ‗affinity‘ for a particular polarity. These probabilities are usually the
result of training linguistic corpora. For example, ‗unforgettable‘ might be assigned a 50% probability of
being indicating a negative affect and a 12.5% probability of being indicating a positive affect and a 37.5%

 ISSN: 2088-8708
Int J Elec & Comp Eng, Vol. 8, No. 5, October 2018 : 4015 – 4022
4016
probability of being indicating a neutral affect as in ‗unforgettable accident‘ or ‗unforgettable party‘. As
stated these probabilities are the result of training corpora; so the bigger and the more general the corpora is
the more reliable and more realistic the probabilities are. This approach outperforms the keywords spotting
approach for giving words realistic polarities and not just plain positive, negative or neutral.
The third approach and the most used one too to create lexicon datasets is using the statistical
methods, such as the Naïve Bayes algorithm, K-Nearest Neighbor or Support vector machines(SVM). It
mainly depends on training a machine learning algorithm with features like words co-occurrence frequencies,
Stylistic features, etc., collected from annotated data and then test the accuracy of the algorithm used on a test
sample from the same data. It is language independent thus avoids ambiguity issues associated with Arabic
[3]. Yet, these methods make classification errors when tested on smaller text units such as clauses as
compared with determining the polarity on the document-level [4]. Gives a quick overview of English SA
research efforts from 2002 up to 2014 that are mostly made using the statistical methods. The authors also
presented some of the available tools and datasets. Furthermore, [5] discussed some of the open issues in the
area of SA including that there is more focus on classing the text into positive and negative only with no
deeper diving in the emotions.
Concept-level sentiment analysis was first introduced by Eric Cambira to classify text based on their
semantics rather than their syntactic through the use of semantic networks like ConceptNet [6] which consists
of nodes representing concepts and connected with edges labeled with common sense 'taken for granted'
information provided by volunteers on the internet. Cambira et al. [7] developed SenticNet, a semantic
resource that uses common sense reasoning techniques along with an emotion categorization model and an
ontology for describing human emotions to infer the polarity of different common sense concepts like
‗beautiful day‘ or ‗feel guilty‘. Each concept is assigned with one float polarity value ∈ [-1,1], followed by
SenticNet2 [8] where more concepts are added allowing a deeper and more multi-faceted analysis of text
while providing a four-dimensional vector (sentic vector) to each concept combined of Pleasantness,
Attention, Sensitivity, and Aptitude and presented as a float value ∈ [-1,1] along with its top-ten affectively
related concepts. Then SenticNet3 [9], which contains both common and common-sense knowledge in order
to boost sentiment analysis tasks such as feature spotting and polarity detection, respectively. Then
SenticNet4 [10], where both verb and noun concepts are linked to primitives so that, for example, concepts
such as attain-knowledge or acquire know-how or acquire-knowledge are generalized as get information. An
addition that allows processing different forms of a concept that otherwise raises a not found error. The idea
of using a generative word was used in other methods too. For example, [11] used synonyms lists for positive
and negative words and mapped the list to one word that already has a polarity value.
To this end, we tested SenticNet4 for the task of polarity detection on a multi-domain Arabic dataset
at the sentence-level and showed results outperforming other Arabic sentiment analysis works that mainly
rely on other approaches. The rest of this paper is organized as follows: the next section reviews research
efforts in the area of Arabic sentiment analysis; followed by a section proposes our framework to detect the
polarity using SenticNet; after which a section discusses the results obtained; finally some concluding
remarks and future work recommendations are presented.
2. LITERATURE REVIEW
This section reviews the contributions to the Arabic sentiment analysis research field. Arabic, which
is the formal language of over than 20 countries around the world and spoken by 300 million native speakers,
is considered under-researched compared to English in the field of sentiment analysis. See Figure 1 that
represents the number of Arabic/English publications per year as presented in [12], [13] and detailed in [14],
[15] respectively. And Arabic is also under-resourced with respect to the amount of data on the internet
knowing that Arabic has scored 4th on the number of web users after English, Chinese and Spanish ranking
the highest growth rate in terms of users with 185 million of users in June 2017 according to
internetwebstats's website: http://www.internetworldstats.com/stats7.htm.
Despite that recent progress, researchers focused on using the statistical methods with a special
focus on supervised machine learning classifiers. They share the same methodology presented in Figure 2
while using different pre-processing and selecting different features. SAMAR [16], a system for Arabic
Subjectivity and Sentiment Analysis, uses Multi-dialectal manually annotated data that covers (Maktoob
chats, tweets, Wikipedia talks and web forums sentences) and does Tokenization, lemmatization and POS
tagging in the Pre- processing step. Then, the system selects syntactic and stylistic features; (Unique: is set
for low frequency words, Polarity Lexicon: checks the presence of positive or negative adjectives, Dialect:
checks the dialects of the text, Gender: checks the gender of text whether it's male, female or unknown, User
ID: checks if the author is a person or an organization and Document ID). They also made experiments with
different combinations of features and the pre-processing tasks while classifying using SVMlight
[17]

Int J Elec & Comp Eng ISSN: 2088-8708 
A Framework for Arabic Concept-Level Sentiment Analysis using SenticNet (Hend G. Hassan)
4017
classifier. Overall results reveal improvements over the baseline performance depending on the training data.
Later on, S. Ibrahim et al. [18] used a manually annotated data and performed normalization and stopwords
removal in the pre-processing step. Then, they selected linguistic and syntactic features like term frequency,
Polar word position, detecting (negation, intensifiers, questions, and supplication) terms along with using the
pattern [adjective + noun] while also classifying using SVMlight
classifier.
Figure 1. Number of Arabic/English publications per year
Figure 2. Supervised machine learning process
3. MODIFYING SENTICNET TO SUIT ARABIC
The framework proposed in [19] is modified to suit the task of Arabic sentences' polarity detection
because Arabic natural language processing tools are trained on and made for modern standard Arabic
(MSA) which is rarely used by internet users' compared to slang Arabic and other Arabic dialects. Figure 3
is an illustration of the proposed framework; Sentences are first decomposed into bi-grams then normalized
and labeled with the part of speech (POS) tags. Then Syntactic patterns like [adjective + noun] are matched
to extract concepts that are translated afterward into English to find a match to in SenticNet.
In order to show the effectiveness of the framework, A multi-domain public dataset is used:
http://bit.ly/1wXue3C, created by ElSahar and El-Beltagy [20], covering Attraction (ATT), Hotels (HTL),
Movies (MOV), Restaurants (RES#1, RES#2) and Products (PROD) reviews. We kept RES#1 as a test
sample. The statistics of the dataset is presented in Table 1 showing the total number of sentences and
concepts we extracted along with the number of positive, negative and mixed sentences number. Those
reviews were rated by their native reviewers then were normalized into the three classes: positive, negative
and mixed following the approach adopted by Pang et al. [21]. The main goal is to conclude the polarity of
each sentence and compare it to the normalized polarity. In order to do that, sentences must be decomposed
into concepts that have a match in SenticNet then the polarity of this match is read. In particular, sentences
are decomposed into bigrams. If a sentence consists of only a bi-gram or a uni-gram, then it is considered a
concept without further analysis. If not the concepts are extracted according to the flowchart proposed in
Figure 3.
0
200
400
600
800
2009 2010 2011 2012 2013 2014 2015
Arabic
English

 ISSN: 2088-8708
4018
Figure 3. Flow chart of the framework to use SenticNet
Before extracting the concepts, sentences must be normalized and structured. To this end, the
following pre-processing steps are followed:
1. Remove elongations.
2. Remove repetitions.
3. Remove punctuations.
4. Remove diacritics.
5. Normalize all Alef forms to ‫ا‬.
6. Normalize ‫ة‬ haa to ‫ي‬.
7. Normalize ِ yaa to ْ.
Table 1. Dataset Statistics
The part of speech (POS) tags of the Normalized text are then generated using MADAMIRA [22], a
shallow syntactic parser that does tokenization, part of speech tagging, and base phrase chunking, and also
combines some of the best aspects of MADA [23] and AMIRA [24]. By reviewing the noun phrases
extracted by MADAMIRA for ATT reviews, we found that around 60% of them are unigrams 20% of which
are a separation of '‫انـ‬ ' the Arabic definite article and about %13 are pronouns separation on the word-level
like ‘‫ي‬ + ‫حصمٕم‬’ (‗his + design‘). Asides to misclassification errors, those 73% are ineffective as concepts.
Thus we used a hand crafted syntactic patterns following the work of ElSahar and El-Beltagy [25] that
extracts slang terms (words/expressions) and transliterated English written in Arabic letters like '‫أَفز‬' that is
transliterated from 'over'. Their work depends on creating a set of lexico-syntactic patterns by using standard
tags like Negator [Neg], person reference [PR], Personal Pronoun [PP], Demonstrative Pronoun [DP],
Intensifier [Ints], Conjunction [Conj], Strong subjective [SS] and the extracted Subjective
Expression is {SE}. For example, "Respectable and very {polite}" would match the pattern: [[SS] [Conj]
{SE} [Ints]], having 'polite' as the extracted term. They created 11 different patterns with a finite set of terms
in each tag and were able to extract 633 unique terms out of 7.5M twitter corpus. In order to be able to match
more patterns, we added to those tags the part of speech tags labeled by MADAMIRA comprising different
patterns as detailed in Table 2. For example [Adjective + [Ints]] would match 'very excellent'.
We also benefited from the fact that the Arabic language has embedded '‫ال‬' in definite nouns and
two consecutive nouns are usually a concept like '‫االخٕزة‬ ‫انسىُاث‬' ('recent years') that can be extracted easily
using the pattern [{} ‫ال‬ }{ ‫ال‬]. Although using syntactic patterns is considered a heuristic method, it extracted
MOV ATT RES#2 PROD RES#1 HTL
#Sentences 1524 2154 2642 4272 8364 15572
#Positive sentences 969 2073 2109 3101 5946 10775
#Negative sentences 384 81 268 863 2418 2647
#Mixed sentences 171 0 265 308 0 2150
#unique concepts 18511 9100 5862 4654 26000 41046

4019
better concepts than using MADAMIRA's noun phrases and verb phrases alone considering the ambiguities
associated with Arabic.
Table 2. Set of Patterns with Matched Examples
Pattern
number
Syntactic pattern
Pattern's
Exception
Example of a
match
Bing Translation Concept
Noun phrases
P1
‫انـ‬{ }‫انـ‬{ }
‫َانـ‬{ }‫انـ‬{ }
‫نال‬{ }‫انـ‬{ }
‫كبنـ‬{ }‫انـ‬{ }
‫فبنـ‬{ }‫انـ‬{ }
‫ببنـ‬{ }‫انـ‬{ }
‫نهـ‬{ }‫انـ‬{ }
Second word
in [ِ‫انذ‬, ّ‫انخ‬ ,
ّ‫انه‬, etc.]
‫األخٕزة‬ ‫انسىُاث‬
ْ‫انقٍز‬ ‫َانُسُاس‬
‫انجدٔدة‬ ‫نألفكبر‬
‫انكُوٕت‬ ُْ‫كبنق‬
ٓ‫انحكُم‬ ‫فبنمسئُل‬
ٓ‫األَن‬ ‫ببندرجت‬
‫انسزٔعت‬ ‫نهُجببث‬
recent years
and obsessive compulsive disorder
to new ideas
as cosmic powers
the government official
in first class
for fast food
recent_year
obsessive_compulsive_disorder
new_idea
cosmic_power
government_official
first_class
fast_food
P2
‫انـ‬{ }Adjective
‫َانـ‬{ }Adjective
‫فبنـ‬{ }Adjective
‫جدٔد‬ ‫انفٕهم‬
ٓ‫عبن‬ ‫َانسعز‬
‫جٕد‬ ‫فبنفٕهم‬
New movie
And the price is high
It is a good movie
new_movie
price_high
good_movie
P3 { }‫ٔه‬{ }‫ٔه‬
[‫انذٔه‬,‫بٕه‬ , ‫مب‬
‫بٕه‬, etc.]
‫مخخهفٕه‬ ‫بمقبسٕه‬ With different sizes different_size
P4 { }‫َن‬{ }‫َن‬ ‫مدربُن‬ ‫انعبمهُن‬ Trained workers Trained
P5 { }‫ان‬{ }‫ان‬
[‫ان‬,‫كبن‬ ,‫َان‬ ,
etc.]
‫فزٔدحبن‬ ‫شخصٕخبن‬ Unique personalities Unique
P6 Adjective + [Ints] ‫ممخبس‬‫جدا‬ Very excellent Excellent
P7 Noun + adjective ‫سزٔعت‬ ‫بحزكبث‬ Quick movements quick_movement
P8
[P_pron] + Adj.
[P_ref] + Adj.
[D_pron] + Adj.
Pattern
followed by
'‫مه‬'
‫محظُظ‬ ‫اوج‬
‫معٕىت‬ ‫وبس‬
‫األفضم‬ ٌُ
Lucky you
Certain people
Is best
Lucky
certain
best
Verb Phrases
P9 Verb + noun ‫َقج‬ ‫َقضٕج‬ I spent time spend_time
Procedure: polarity detection
Input:
English translation of the patterns
Output:
Polarity, Pleasantness, Attention, Sensitivity, Aptitude, Semantics
Begin:
For each pattern in the sentence:
Remove punctuations
Remove stopwords
lemmatize the nouns
If the first word's tag is verb
Lemmatize the verb
Search for a match in SenticNet
End
Sum the polarities of each sentence
If classifying into three classes
If (sum >= 1)
Positive
Else if (sum <=- 1)
Negative
Else
Mixed
End
Else If classifying into two classes
If (sum >= 0)
Positive
Else if (sum <0)
Negative
End
End
End
Figure 4. Pseudo code for polarity detection
Next, we used Microsoft Bing translator to translate the matches to English. Having English on the
output side of the machine translation system and not translating concepts from SenticNet into Arabic avoids
the ambiguity of different dialect candidates and different sentence structures; Arabic is one of the languages

 ISSN: 2088-8708
4020
that has multiple sentence forms [subject-verb-object (SVO), verb-subject-object (VSO), verb-object-subject
(VOS) asides to the possibility of having a correct sentence dropping a verb or copula].
At last, the translated extracted concepts follow the steps presented in Figure 4 to match the same
form of SenticNet's concepts where nouns are singular and verbs are lemmatized. If a match is found in
SenticNet then the polarity value is read. If not a search for a match for the first word of the concept is done
as it is usually an adjective for example '‫رائع‬ ‫'فىدق‬ (wonderful hotel) would extract 'wonderful' if the whole
concept 'wonderful_hotel' is not found in SenticNet.
4. RESULTS AND DISCUSSION
In order to properly evaluate the performance of the proposed framework, we used the leading
measuring methods in the NLP Classification process: precision, recall, f-measure and accuracy that are
shown by values in table3 respectively for each dataset in case of 2-class classification problem (positive and
negative) and 3-class classification problem (positive, negative and mixed).We highlighted the datasets with
the best scores revealing that the 2-class classification problem has better results than the 3-class
classification problem for the same dataset. The same result was also obtained in [20]. And to show the
difference between our method and existing ones, we compared these results with the ones obtained in [20] in
which they used the same dataset we used but while using the statistical methods (See Table 4). The reported
average accuracy in Table 3 is the average of all accuracies reported after using different lexicon based
features. Their reported accuracy is a result of training 80% of the data with a machine learning classifier and
calculating accuracy on a 20% test sample from the same data with the classifier. Furthermore, best accuracy
score in the 2-class classification happens for the ATT dataset. This could be explained by the fact that it has
more concepts extracted as compared to RES#2 that has more sentences but fewer concepts and for which it
scored the second best accuracy.
Table 3. The Value of Precision, Recall, F-measure and accuracy respectively for each Dataset
P R F1 Acc. ElSahar's work average accuracy
3-class
ATT There are no mixed polarity reviews in the ATT dataset Not mentioned
PROD .57 .45 .49 .45 0.51
MOV .51 .62 .54 .62 0.47
RES#2 .72 .72 .71 .72 0.57
HTL .55 .62 .58 .62 0.64
2-class
ATT .96 .86 .91 .89 Not mentioned
PROD .78 .91 .84 .73 0.74
MOV .73 .91 .81 .70 0.69
RES#2 .91 .91 .91 .85 0.81
HTL .81 .87 .84 .73 0.85
Table 4. Comparison between our Results and ElSahar's Results
Average accuracy
Dataset lexicon
3-class 2-class
ElSahar's work .56 .77 2k un normalized entries
Proposed framework .60 .75 96k unique entries
Figure 5 is a boxplot of the number of words and concepts for each dataset and it shows that ATT
has more words in the sentence than RES#2 causing more concepts. Although MOV dataset has relatively
more concepts, it scored last. That can be explained by the fact that it has the longest review length as it has
2530 words in one of the reviews. The 3-class classification problem has the same ranking order except for
the PROD dataset that has fallen behind as it has uni-gram reviews.
On the other hand, ElSahar 's lexicon [20] has around 2000 entries (uni-grams and bi-grams) that are
not normalized nor lemmatized;'‫أوصح‬' (I recommend) , '‫بشزاء‬ ‫اوصح‬' (I recommend to buy) ,'ً‫ب‬ ‫اوصح‬ '
(recommend it) ,' ‫بٍب‬ ‫اوصح‬ ' (I recommend it/female pronoun),' ‫اوصحكم‬ ' (I recommend you ) are all entries and
all of them has the same lemma 'recommend' while we were able to extract around 69 k unique entries after
removing redundancy from the different datasets. Furthermore, we used a test sample from the dataset
(RES#1) in order to validate our lexicon by following the same steps in the framework while skipping the
translation step as shown in Figure 6. We were able to match 68% of the concepts extracted from RES#1 in
the lexicon. The accuracy obtained was 70% and the precision was 70% with a recall of 100% and an F-
measure of 82%.

4021
Figure 5. Boxplot's section showing the number of words and concepts for each dataset
Figure 6. Flow chart for the testing framework of the lexicon dataset
5. CONCLUSION AND FUTURE WORK
A novel framework for concept-level sentiment analysis was introduced to detect the polarity of
Arabic sentences using Senticnet. The framework is created so that it can handle ambiguity issues associated
with Arabic including the fact that slang Arabic lacks syntactic rules and tools to deal with and it also doesn't
include using any machine learning algorithm.The framework was tested on a multi-domain dataset covering
public reviews scrapped from the internet. The results showed promising performance as the accuracy
reached 89%. it also outperformed other research works in terms of detecting the polarity of a sentence
without having prior annotated data. In the future, we plan on Handling Polarity inversion terms such as
negations and also defining the scope of each negating term along the sentence.
REFERENCES
[1] E. Cambira, et al., ―The CLSA Model: A Novel Framework for Concept-Level Sentiment Analysis‖, 16th
international Conference,CICLing 2015, Part II, pp. 3-22, 2015.
[2] E. Cambira, ―An Introduction to Concept-Level Sentiment Analysis‖, The 12th Mexican International Conference
on Artificial Intelligence, Part II, pp. 478-483, 2013.
[3] N. Y. Habash, ―Introduction to Arabic Natural Language Processing‖, Synthesis Lectures on Human Language
Technologies 3.1, 2010.
[4] A. Dhokrat, et al., ―Review on Opinion Mining for Fully Fledged System‖, Indonesian Journal of Electrical
Engineering and Informatics (IJEEI), vol. 4, no. 2, pp. 141-148, 2016.
[5] P. Kumar and S. Nandagopalan, ―Insights to Problems, Research Trend and Progress in Techniques of Sentiment
Analysis‖, International Journal of Electrical and Computer Engineering (IJECE), vol. 7, no. 5, pp. 2818-2822,
2017.
[6] H. Liu and P. Singh, ―ConceptNet — A Practical Common Sense Reasoning Tool-kit‖, BT Technology Journal,
vol. 22, no. 4, pp. 211-226, 2004.
[7] E. Cambria, et al., ―SenticNet: A Publicly Available Semantic Resource for Opinion Mining‖, AAAI Fall
Symposium: Commonsense Knowledge, vol. 10, 2010.
[8] E. Cambria, et al., ―SenticNet 2: A Semantic and Affective Resource for Opinion Mining and Sentiment Analysis‖,
FLAIRS conference, pp. 202-207, 2012.
[9] E. Cambria, et al., ―SenticNet 3: A Common and Common-Sense Knowledge Base for Cognition-Driven Sentiment
Analysis‖, Twenty-eighth AAAI conference on artificial intelligence, 2014.

 ISSN: 2088-8708
4022
[10] E. Cambria, et al., ―SenticNet 4: A Semantic Resource for Sentiment Analysis based on Conceptual Primitives‖,
COLING, pp. 2666-2677, 2016.
[11] P. Arora, et al., ―An Approach for Big Data to Evolve the Auspicious Information from Cross-Domains‖,
International Journal of Electrical and Computer Engineering (IJECE), vol. 7, no. 2,pp. 967-974, 2017.
[12] A. AlOwisheq, et. al., ―Arabic Sentiment Analysis Resources: a Survey‖, Social Computing and Social Media: 8th
International Conference, pp. 267-278, 2016.
[13] K. Dashtipour, et al., ―Multilingual Sentiment Analysis: State of the Art and Independent Comparison of
Techniques Cognitive Computation‖, vol. 8, no. 4, pp. 757-771, 2016.
[14] Al-Twairesh, et al., ―Subjectivity and Sentiment Analysis of Arabic: Trends and Challenges‖, 11th International
Conference on Computer Systems and Applications (AICCSA), IEEE/ACS, pp. 148-155, 2014.
[15] K. Ravi and V. Ravi, ―A Survey on opinion mining and sentiment Analysis: Tasks, Approaches and Applications‖,
Knowledge-Based Systems, vol. 89, pp. 14-46, 2015.
[16] M. A. Mageed, et al., ―SAMAR: A System for Subjectivity and Sentiment Analysis of Arabic Social Media‖,
Computer Speech & Language, vol/issue: 28(1), pp. 20-37, 2014.
[17] T. Joachims, ―Svmlight: Support vector machine‖, Cornell University, 2008.
[18] H. Ibrahim, et al., ―Sentiment Analysis for Modern Standard Arabic and Colloquial‖, International Journal on
Natural Language Computing (IJNLC), vol. 4, no. 2, pp. 967-974, 2015.
[19] E. Cambria and A. Hussain, ―Sentic Computing A Common-Sense-Based Framework for Concept-Level Sentiment
Analysis‖, Springer, vol. 1, 2015.
[20] H. ElSahar and S. R. El-Beltagy, ―Building Large Arabic Multi-domain Resources for Sentiment Analysis‖, 16th
international Conference,CICLing 2015, Part II, pp. 23-34, 2015.
[21] B. Pang, et al., ―Thumbs up? Sentiment Classification using Machine Learning Techniques‖, Proceedings of the
ACL-02 Conference on Empirical Methods in Natural Language Processing, vol. 10, pp. 79-86, 2002.
[22] A. Pasha, et. al., ―MADAMIRA: A Fast, Comprehensive Tool for Morphological Analysis and Disambiguation of
Arabic‖, LREC, 2014.
[23] N. Habash, et. al., ―MADA+TOKAN:A Toolkit for Arabic Tokenization, Diacritization, Morphological
Disambiguation, POS Tagging, Stemming and Lemmatization‖, Proceedings of the 2nd International Conference
on Arabic Language Resources and Tools(MEDAR), Cairo, Egypt, pp. 102-10.
[24] M. Diab, ―Second Generation AMIRA Tools for Arabic Processing Fast and Robust Tokenization, POS tagging,
and Base Phrase Chunking‖, Proceedings of KONVENS, pp. 39-52, 2012.
[25] H. ElSahar and S. R. El-Beltagy, ―A Fully Automated Approach for Arabic Slang Lexicon Extraction from
Microblogs‖, 15th International Conference, CICLing 2014, Part I, pp. 79-91, 2014.

A Framework for Arabic Concept-Level Sentiment Analysis using SenticNet

More Related Content

What's hot (20)

Similar to A Framework for Arabic Concept-Level Sentiment Analysis using SenticNet (20)

More from IJECEIAES (20)

Recently uploaded (20)

A Framework for Arabic Concept-Level Sentiment Analysis using SenticNet