Hybrid Sentiment Analyser for Arabic Tweets using R

Sarah Alhumoud, Tarfa Albuhairi and Wejdan Alohaideb

College of Computer and Information Science, Al-Imam Muhammad Ibn Saud Islamic University, Riyadh, Saudi Arabia

Keywords: Sentiment Analysis, Data Mining, Machine Learning, Supervised Approach, Hybrid Learning.

Abstract: Harvesting meaning out of massively increasing data could be of great value for organizations. Twitter is

one of the biggest public and freely available data sources. This paper presents a Hybrid learning

implementation to sentiment analysis combining lexicon and supervised approaches. Analysing Arabic,

Saudi dialect Twitter tweets to extract sentiments toward a specific topic. This was done using a dataset

consisting of 3000 tweets collected in three domains. The obtained results confirm the superiority of the

hybrid learning approach over the supervised and unsupervised approaches.

1 INTRODUCTION

Making one’s voice reachable instantaneously to the

globe was a sort of a dream before. In some parts of

the world where opinion freedom is in its lower

indexes there are barley any channels to vent and

express opinions. Now, with the emergence of social

networks, opinions could be shared with the globe in

real time. The power of one video on YouTube, a

post on Facebook or a tweet on Twitter can change a

country’s fate and can be an effective means of

marketing (Dinh et al., 2014). A study done by

(Storck, 2011) has displayed the role of social media

in the Egyptian revolution and how it helped the

revolutionaries to achieve their goals. It has been

known now by decision makers in political and

economic fields that they need to hear the client’s

voice and immediate opinions and to study their

behaviour for them to gain further insight and move

wisely. A study by (Dinh et al., 2014) has shown the

effect of viral marketing on social networks and how

to find the optimal seeding for the aids in social

networks. An example of a mining application to

social networks seeking value and meaning is IBM

customer analytic tool which can measures customer

sentiment through surveying data in social networks

to make a company more aware of its customers,

who they are and what they want (IBM, 2011).

Social networks' applications offer means by

which people build cyber social bonds based on their

opinions, orientations, speciality, or hubbies and

preferences. Considering the low freedom of

expression in the Arab region by classical means,

social networks have been an important ventilation

mechanism. According to freedom house 2015

report, the Arab region is mostly ranked as low in

civil liberties and political rights (Freedom House,

2015). In 2014, the number of Arabic internet users

was 135 million, with more than 71 million of them

are social networks active users (Mohammed Bin

Rashid School of Government, 2014). Figure 1(a)

shows the Arab usage of social networks in 2014.

Moreover, a report was released in 2014 stating that

the number of Arab Twitter active users is about 6

million while 2.4 million of them are from Saudi

Arabia which makes it the highest Arab country in

Twitter usage as shown in figure 1(b) (Arab Social

Media Report, 2014). In Saudi Arabia, 60% of social

networks’ users were using Twitter in the year of

2013 (Reyaee and Ahmed, 2015). Having this

extensive use of Twitter by the Saudi tweeters

implies streaming a large amount of data that could

be a rich and an invaluable source for analysis and

study.

Sentiment Analysis (SA) is considered one of the

text mining tasks (Han, Kamber and Pei, 2000) and

one of the Natural Language Processing (NLP)

concepts (Liu, 2012). SA is mainly the process of

classifying text into two classes positive and

negative to conclude the writer’s orientation towards

a certain topic or subject (Liu, 2012). After

reviewing the literature, two approaches to

implement SA have been concluded. The first is

supervised approach or corpus-based

approach(Abdulla et al., 2013). The second approach

to SA is the unsupervised approach also called

lexicon-based approach (Liu, 2012) and (Alhumoud

et al., 2015).

Alhumoud, S., Albuhairi, T. and Alohaideb, W..

Hybrid Sentiment Analyser for Arabic Tweets using R.

In Proceedings of the 7th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2015) - Volume 1: KDIR, pages 417-424

ISBN: 978-989-758-158-8

417

Figure 1(a): Arab internet usage in 2014.

Figure 1(b): Arab Twitter usage in 2014.

One of the data mining tools that is used to

implement SA is R programming language (R-

project, 2015) R is an open source programming

language that can perform data mining main tasks

such as clustering, and classification (Kosorus,

Honigl and Kung, 2011).

This research presents an implementation of SA

using R programming language applying the hybrid

learning approach and comparing it to the supervised

and unsupervised learning approaches. The

methodology and results of each approach will be

discussed later in this paper.

This paper is organised as following: related

work section where a list of similar work presented

and discussed. Then, the methodology section that

describes the method that was followed to

implement the learning schemes. After that, the

results and discussions’ section, presenting results of

each approach. Finally, the conclusion where the

paper is wrapped up pointing out the marked

findings.

2 RELATED WORK

Based on a survey that was published in 2014 by

(Medhat, Hassan and Korashy, 2014) there are two

main approaches for sentiment analysis. Those are

machine learning and lexicon based approaches.

Supervised learning (also known as corpus-based

approach) falls under machine learning approach.

There are several data mining algorithms that can be

used to implement the supervised learning such as

such as Support Vector Machine (SVM), Naïve

Bayes (NB), Decision Tree (D-Tree) and K-Nearest

Neighbors (KNN). The supervised learning consists

of five sequenced stages based on multiple papers

that implemented SA supervised learning such as

(Shoukry and Rafea, 2012) and (Abdulla et al.,

2013). These stages are: building the dataset,

building the classifier (model), training the

classifier, testing the classifier and finally using the

classifier. Building the dataset is the first stage,

creating labelled data and dividing it into training

dataset and testing dataset. The second stage is

building the classifier model where one of the data

mining algorithms is chosen as a learning approach

for the classifier. After that the classifier needs to be

trained using the training dataset enabling it to

classify new data. Once the classifier has been

trained its accuracy is measured by testing it using

the testing dataset. Where the classifier hide the

labels and tries to classify the data based on the

experience it gained in the training stage then it

compares the classification result with the actual

values. Finally, a new dataset with no labels is fed

into the classifier to find its ability to classify based

solely on previous learning.

Lexicon based approach, also known as

unsupervised approach is a way where the classifier

rely on number of lexicons helping it to classify the

text. A lexicon is list of words or phrases, idioms or

adjectives that hold a sentimental meaning. There

are several ways to construct a lexicon, one of them

is dictionary based. It starts with seeds of

sentimental words then increasing the size of the

dictionary by adding the synonym and antonyms

(Medhat, Hassan and Korashy, 2014). Once a

lexicon is completed it is ready for the classifier. An

unsupervised classifier is an algorithm that performs

matching between lexicon words and the words in

the text. It usually treats words (in the lexicon and

text) as numbers. Where positive words represented

as positive integers and negative words as negative

integers. Additionally, using certain mathematical

processes it assigns the polarity of whole text as: +1,

-1 or 0 for positive, negative or neutral, respectively

(Shoukry and Rafea, 2012). There is no limitation

for the number of lexicons that can be used for

classification such as the work proposed by

(Ibrahim, Abdou and Gheith, 2015) where they used

two lexicons one holds sentimental words and the

other for phrases and idioms.

Other

135 M

Social

Media

71 M

Arab

5.7 M

Saudi

2.4 M

KDIR 2015 - 7th International Conference on Knowledge Discovery and Information Retrieval

418

The tools that are used for SA purposes are wide

such as R programming language. The interest in R

language for data mining is growing rapidly (KD

Nuggets, 2013) (Magoulas and King, 2014). This

increased interest in R is noticeable in both industry

and academia. A key advantage of R programming

language other than being an open source is data

structures, visualization and packages. R provides a

wide range of data structures that can represent data

in understandable forms. Also, R provides the ability

to transform results into graphs which makes it

easier to understand. R has lots of packages that are

specialized to different tasks and domains. (Jovi,

Brkiand and Bogunovi, 2014). Authors (Kumar et

al., 2011) performed clustering using R to

implement a framework that provides a platform for

developing data mining application. They used k-

means and fuzzy k-means algorithms where fuzzy k-

means got better results. While (Adamov, 2014)

have used R language to pre-process http server logs

for web usage mining. Moreover, the Authors (Nisa,

Andrianto and Mardhiyyah, 2014) tried to solve the

problem of forest fire by applying clustering

function using R language.

R language is used to perform classification that

solves problems of different domains. Work by

(Keka and Hamiti, 2013) shows that they applied

data mining to show electricity consumption over

time. Electricity loads classification was based on

two factors: seasons (summer and winter) and the

usage intervals during the day. They applied

unsupervised approach to achieve the loads

classification. Another interesting work done by

(Tsatsoulis and Hofmann, 2014) in that they tried to

use data mining to classify albums of the rock

musician Tom Waits into two classes. The first class

is for albums that were produced before 1983

(period A) and second class for albums that were

produced after 1983 (period B). They used word

cloud package in R in order to help them to visualize

the two classes. Their results using RTextTools

package show that Maximum Entropy was able to

classify the albums correctly with 95% of accuracy.

A work presented by (Hosch, 2014) using R

programming language and nine algorithms. The

aim was to build a classifier that is able to classify

open source projects into functional categories.

Their experiments gave the classifier an accuracy of

81%.

R language has been used for SA purposes by

many researchers as SA is one of the applications to

classification (Ofek, 2015).The authors of (Fiaidhi et

al., 2012) used TwitteR package to implement

twitter client and used the searchTwitter() function

to collect tweets from twitter. They applied an

unsupervised approach. For the supervised approach

authors (Horakova, 2015) have propose SA machine

learning for Czech language. They hired RTextTools

to build classifiers with five algorithms. They

collected a dataset of 7000 users review about

electronic devices. Also, a pre-processing model was

applied. For filtering, they have used tagging,

tokenization, stemming and lemmatization. The

result of the experiments shows that SVM is among

the best classifiers with 73% while NB is the worst

with 51% accuracies respectively. Their results also

show that all the classifiers have improved their

accuracy when increasing the size of the dataset

except for NB. They also, have experimented the

classification on raw data and the results show a

decrease in the accuracy. Table 1 compares between

papers that used R to apply classification function,

how they apply it, and to solve what problem.

3 METHODOLOGY

For the convenience of R language, the hybrid

learning approach was implemented using R. R

provided two packages that facilitates the

implementation of SA, those are RTextTools and tm

packages. RTextTools was designed to make

machine learning accessible (Jurka et al., 2015).

Whereas tm was designed to apply text mining

techniques (Feinerer and Hornik, 2015). SA

implementation was done using the SVM data

mining algorithm.

In general, the hybrid learning approach contains

five main stages: building the dataset, building the

classifier (model), training the classifier, evaluating

the classifier, and using the classifier to get overall

sentiment of a new dataset as shown in Figure 2. The

main difference between supervised and hybrid

learning is in building the training dataset, this will

be explained in the next section.

3.1 Building the Training Dataset

The training dataset is built from rows of single

sentimental words and their labels. A labelled word

tagged positive or negative according to the word

sentiment. The training dataset contains 3690

sentimental words, 1370 words were positive and

remaining 2320 words are negative. The training

dataset contains 1000 MSA sentimental words are

found in (Arabic MPQA subjective lexicon and

Arabic opinion holder corpus, 2012) and 2690 Saudi

dialect sentimental words were added manually with

Hybrid Sentiment Analyser for Arabic Tweets using R

419

Table 1: Comparison of classification approaches in related work.

Paper citation Application Data source Dataset size Pre-processing Filtering Results and approach

(Tsatsoulis

and Hofmann,

2014)

Classify albums

based on publishing

periods by analysing

song’s lyrics.

Lyrics of rock

musician Tom

Waits

20 albums

 Converting to

lowercase

 Removing

punctuation

 Tokenizing

 Removing

stop words

 Supervised

 Maximum Entropy 95%.

(Hosch, 2014)

Classify open source

projects into

functional categories.

Ohloh

database

Not defined -  Tagging

 Supervised

 Used 8 algorithms

 Accuracy 81%

(Keka and

Hamiti,

20113)

Electricity

consumption over

time.

Electrical

working

stations

2,976

observations

- -

(Horakova,

2015)

SA for Czech

language.

Czech portal

heureka.cz

7,000 user

reviews

Text pre-

processing

module.

 Tagging

 Tokenization

 Stemming

 Lemmatization

 Supervised

 Used 8 algorithms

 Best accuracy:

 Maximum Entropy

 Random Forests

 Increasing dataset size rise accu-

racy in all algorithms except NB.

 Effect of no filtering and pre-

processing is decreasing accuracy.

(Fiaidhiet al.,

2012)

SA on tweets Twitter 1500 tweets -  TF-IDF  Unsupervised

the help of 1000 sports tweets collected form

Twitter.

In R language, the dataset need to be represented

in a specific format such as Comma Separated Value

(CSV). Create_matrix() function was used to

generate the document term matrix. Several pre-

processing options from the tm package are

available, including stripping whitespace, removing

sparse terms and setting number of minimum

frequency of sentimental words in the whole

document. In addition, word stemming, and stop

word removal for several languages. However,

currently those tools do not support Arabic

language. Stop words were removed manually using

simple algorithm with stop words’ list. The hybrid

learning approach could be implanted on languages

other than Arabic, such as English. For the English

language it is even more convenient and easy to use

the full features of R SA packages.

In the supervised approach, each line contains

one instance, a tweet. This instance has several

words that do not affect the sentiment but causes

confusion in the classifier, hence, decreasing the

accuracy. If these words are removed from the

instance, only sentimental words will remaining

which is similar to the lexicon in the unsupervised

approach. Using sentimental lexicon as a training

dataset minimises the classifier confusion.

Therefore, the hybrid learning incorporates the

advantages of a data mining algorithm in supervised

approach and lexicon based approach in

unsupervised approach to better teach the classifier.

Figure 2: Classification stages of hybrid approach.

The difference between supervised and hybrid

learning in building the training dataset reflected a

remarkable difference in the results as well.

The hybrid learning also avoids the pre-

processing cost associated with the supervised

approach. Building a training dataset in the

supervised learning requires several steps, first,

collecting n instances, second pre-processing them

with O(n) time complexity, third normalizing them

with O(n) time complexity as well. While the hybrid

learning approach does not require the second nor

third steps.

KDIR 2015 - 7th International Conference on Knowledge Discovery and Information Retrieval

420

3.2 Building the Classifier

Second step to the classification is building the

classifier. SVM data mining algorithm was used to

build the classifier. This algorithm works efficiently

in text classification and it showed superior

performance in previous related studies (Khasawneh

et al., 2013), (Shoukry and Rafea, 2012) and

(Abdulla et al., 2013).

3.3 Training the Classifier

Training the classifier is the third step in

classification approaches. SVM classifier used linear

kernel, unigram case and the minimum term

frequency which is equal to one. Unigram means

reading the dataset word by word, it is one of the n-

gram cases which splits a string into n-grams with

min and max gram.

3.4 Evaluating the Classifier

Evaluating the classifier is the fourth step in the

classification which uses the training dataset as

testing dataset and the produced classifier to expect

the sentiment of each tweet. The result of the

evaluation is measured by computing the precision,

and recall. The precision and recall equations are

presented below:

Precision = TP / (TP + FP)

(1)

Recall = TP / (TP + FN)

(2)

Where TP, FP, TN, and FN are true positive,

false positive, true negative and false negative,

respectively.

In precision and recall, the highest level of

performance is equal to one and the lowest is zero

(Sokolova et al., 2006). If precision and recall was

more than 0.5, the classifier is accepted to classify

new dataset.

3.5 using the Classifier

Fifth step comprises the classification of a new

dataset by using the produced classifier to expect

overall sentiment polarity of each tweet in the

dataset. The tweets must be normalized before

building the dataset. In addition, the dataset should

be using the same format of the built container.

Classification accuracy was measured by computing

the number of correct classified tweets.

4 RESULTS AND DISCUSSIONS

This research presents the implementation of the

hybrid learning approach compared to the

supervised and unsupervised learning approaches.

The same dataset is used for all three

implementations. Both hybrid learning and

supervised learning used SVM. Supervised classifier

trained on sports domain using 1000 tweets. The

unsupervised approach has two words’ dictionaries,

positive and negative. Positive words were 1370

words, while 2320 words are negative. These words

are the same words used to train the hybrid

classifier.

The hybrid learning classifier precision and

recall were 0.960 and 0.970, respectively. Table 2

shows a comparison between the accuracy of the

three classifiers hybrid learning, supervised and

unsupervised, denoted by "Hyb", "Sup" and "Unsup"

consecutively. These classifiers were classifying

new dataset in different sizes.

Table 2: Accuracy of classifiers classifying sports domain.

Dataset

Size

Hyb Sup Unsup

100

91.50% 90.00% 86.30%

250

90.60% 81.50% 80.70%

500

89.90% 80.80% 78.60%

1000

89.30% 84.30% 81.30%

Average

90.30% 84.20% 81.70%

Figure 3: Accuracy for hybrid learning, supervised and

unsupervised approaches in three different domains.

The results of classifying several datasets in

sports domain proved superiority of the hybrid

learning classifier over both supervised and

unsupervised approaches scoring an improvement in

100

Soprts Social Political

Accuracy percentage

Domains

Hybrid Learning Supervised Unsupervised

Hybrid Sentiment Analyser for Arabic Tweets using R

421

accuracy of 6% and 8% over supervised and

unsupervised approaches respectively.

Figure 3 shows the comparison in accuracy of

the three learning approaches in three domains,

sports, social and political, each having 1000, 500

and 500 tweets respectively. The classifiers were

trained using the sports domain. Testing the

classifiers using new domains will exercise the

learning scalability of the classifiers.

The proposed hybrid classifier scored an increase

in accuracy of 7%, 15% and 19% in sports social

and political domains respectively over the

supervised approach.

Additionally, a similar enhancement was

noticeable with the hybrid learning over the

unsupervised approach scoring 9%, 12% and 14% in

sports social and political domains respectively. The

results show that when using the same SVM

classifiers to classify new datasets in new domains,

the hybrid learning approach achieve better accuracy

than supervised and unsupervised. Proving a better

scalability in different domains.

Moreover, increasing the training dataset size in

hybrid learning is easier than the supervised

approach as it does not require the pre-processing

steps as in the supervised approach.

Enhancing the classifier needs some insight on

the reasons that might lead to the incorrect

classification. The following are examples of

situations that may lead to classifier confusion or

error:

• A tweet that has a negation word and this

inverts the tweet polarity.

•

A tweet having two opposite sentiments and

could be classified to the wrong polarity.

•

A tweet that is having an ambiguous or

unknown sentiment to the classifier.

•

The sentiment of a tweet is specified by a non-

sentimental word that could not be added to the

training set

•

Words that have two different sentiments in two

different domains

The hybrid learning approach’s accuracy can be

improved by increasing the size of the training

dataset, and by using a words’ stemmer.

5 CONCLUSIONS

This paper presented the new hybrid learning

approach for Arabic SA in Twitter examining the

classification of randomly collect tweets in three

domains. The results confirm that the hybrid

learning approach has better accuracy than both

supervised and unsupervised approaches. The hybrid

learning approach scored an enhancement in

accuracy of 7%, 15% and 19% in sports, social and

political domains over the supervised approach and a

similar enhancement over the unsupervised

approach. Moreover, the hybrid learning approach

proves better scalability when applied to new

different domains over the supervised and the

unsupervised approaches. Another convenient

feature of the hybrid approach is avoiding the

overhead associated with the supervised pre-

processing approach.

REFERENCES

Abdulla, N. Ahmed, N. Shehab, M. & Al-Ayyoub, M.

(2013) Arabic Sentiment Analysis: Lexicon-Based and

Corpus-Based. Proceedings of the IEEE Jordan

Conference on Applied Electrical Engineering and

Computing Technologies (AEECT). Amman, pp. 1- 6.

Adamov, A. (2014) Data Mining and Analysis in Depth.

Case Study of QAFQAZ University HTTP Server Log

Analysis. Proceedings of the IEEE 8th International

Conference on Application of Information and

Communication Technologies (AICT), Astana, pp.1-4.

Alhumoud, S. Altuwaijri, M. Albuhairi, T. & Alohaideb,

W. (2015) Survey on Arabic Sentiment Analysis in

Twitter. Proceedings of the International Conference

on Computer Science and Information Technology

(ICCSIT), Paris, pp. 364 – 368.

Arab Social Media. (2014) Twitter in the Arab Region,

Dubai School of Government, [Online] Available

from: https://shar.es/129INW. [Accessed: 16th June

2015].

Dinh, N. Huiyuan, Z. Nguyen, .T. & Thai, T. (2014) Cost-

Effective Viral Marketing for Time-Critical

Campaigns in Large-Scale Social Networks,

IEEE/ACM Transactions on Networking, 22, pp.2001-

2011

EthraArab. (2015)



Stop Words

 

, [Online]

Available from: http://wp.me/p5VrNb-5o. [Accessed:

8th July 2015].

Feinerer, I. & Hornik, K. (2015) Package ‘tm’. [Online]

Available from: http://cran.r-

project.org/web/packages/tm/tm.pdf. [Accessed: 17th

June 2015].

Fiaidhi, J.Mohammed, O.Mohammed, S.Fong, S.& Kim,

T. (2012) Opinion Mining Over Twitterspace:

Classifying Tweets Programmatically Using The R

Approach. Proceedings of the Seventh International

Conference on Digital Information Management,

Macau, pp. 313 - 319.

Freedom House, (2015) Freedom in the World, [Online]

Available from: https://freedomhouse.org/report-

KDIR 2015 - 7th International Conference on Knowledge Discovery and Information Retrieval

422

types/freedom-world#.VZcL2PkW6wN. [Accessed:

4th July 2015].

Han, J. Kamber, M. & Pei, J. (2000) Data Mining:

Concepts and Techniques. Morgan Kaufmann.

Horakova, M. (2015) Sentiment Analysis Tool Using

Machine Learning. Proceedings of the 2nd Global

Conference on Computer Science, Software, Networks

and Engineering. Turkey, pp. 192-204.

HOSCH, L. (2014) Using Machine Learning To Classify

Open Source Projects. Department of Computer

Science, Bachelor Thesis, University of Friedrich-

Alexander. [Online] Available from:

http://dirkriehle.com/uploads/byhand/theses/2014/hoes

ch_2014_arbeit.pdf. [Accessed: 16th June 2015].

IBM. (2011) Customer Analytics Pay off, IBM. [Online]

Available from: http://www-01.ibm.com/software/

analytics/rte/an/customer-analytics/. [Accessed: 1st

September 2015].

Ibrahim, H. Abdou, S. & gheith, M. (2015) Sentiment

Analysis for Modern Standard Arabic and Colloquial,

International Journal on Natural Language Computing

(4). [Online] Available from: arxiv.org/ftp/arxiv/

papers/1505/1505.03105.pdf. [Accessed: 27th Aug

2015].

Jovi, A. Brki, K. & Bogunovi, N. (2014) An Overview

Of Free Software Tools For General Data Mining.

Proceedings of the International Convention on

Information and Communication Technology,

Electronics and Microelectronics (MIPRO), Opatija,

pp. 1112 - 1117.

Jurka, T. Collingwood, L. Boydstun, A. Grossman, E. &

Atteveldt, W. (2015) Package ‘RTextTools’. [Online]

Available from: http://cran.r-project.org/web/

packages/RTextTools/RTextTools.pdf. [Accessed:

17th June 2015].

KD Nuggets. (2013) Top Languages For Analytics, Data

Mining, Data Science, [Online] Available from:

http://www.kdnuggets.com/2013/08/languages-for-

analytics-data-mining-data-science.html. [Accessed:

16th June 2015].

Keka, I. & Hamiti, M. (2013) Load Profile Analyses

Using R Language. Proceedings of the International

Conference on Information Technology Interfaces

(ITI), Cavtat, pp. 245 - 250.

Khasawneh, R. Wahsheh, H. Al Kabi M. & Aismadi, I.

(2013) Sentiment Analysis of Arabic Social Media

Content: A Comparative Study. Proceedings ofthe8th

International Conference for Internet Technology and

Secured Transactions (ICITST). London, pp.101-106.

Kosorus, H. Honigl, J. & Kung, J. (2011) Using R, WEKA

and RapidMiner in Time Series Analysis of Sensor

Data for Structural Health Monitoring. Proceedings of

the 2nd International Workshop on Database and

Expert Systems Applications, Toulouse, pp. 306 - 310.

Kumar, P. Ozisikyilmaz, B. Liao, W. Memik, G. &

Choudhary A. (2011) High Performance Data Mining

Using R on Heterogeneous Platforms. Proceedings of

the 2011 IEEE International Symposium on Parallel

and Distributed Processing Workshops and PHD

Forum (IPDPSW), Shanghai, pp. 1720-1729.

Liu, B. (2012) Sentiment Analysis and Opinion Mining.

Morgan & Claypool.

Magoulas, R. & King, J. (2014) 2013 Data Science Salary

Survey Tools, Trends, What Pays (and What Doesn't)

For Data Professionals. O'Reilly, [Online] Available

from: http://www.oreilly.com/data/free/files/strata

survey.pdf. [Accessed: 16th June 2015].

Medhat, W, Hassan, A. Korashy H. (2014) Sentiment

analysis algorithms and applications: A survey, Ain

Shams Engineering Journal (5). P. 1093–1113.

[Online] Available from: http://www.sciencedirect.

com/science/article/pii/S2090447914000550.[Accesse

d: 27th Aug 2015].

Mohammed Bin Rashid School of Government. (2014)

The Arab World Online 2014: Trends in Internet and

Mobile Usage in the Arab Region, Mohammed Bin

Rashid School of Government. [Online] Available

from: http://www.mbrsg.ae/getattachment/ff70c2c5-

0fce-405d-b23f-93c198d4ca44/The-Arab-World-

Online-2014-Trends-in-Internet-and.aspxl. [Accessed:

2th July 2015].

Nisa, K. Andrianto, H. & Mardhiyyah, R. (2014) Hotspot

Clustering Using DBSCAN Algorithm and Shiny Web

Framework. Proceedings of the International

Conference on Advanced Computer Science and

Information Systems (ICACSIS), Jakarta, pp. 129-132.

NLP for Arabic. (2012) Arabic MPQA Subjective Lexicon

& Arabic Opinion Holder Corpus. [Online] Available

from: http://nlp4arabic.blogspot.com/2012/05/arabic-

mpqa-subjective-lexicon-arabic.html. [Accessed: 17th

June 2015].

Ofek, N. Rokach, L. Caragea, C. & Yen, J. (2015) The

Importance of Pronouns to Sentiment Analysis: Online

Cancer Survivor Network Case Study, Proceedings of

the 24th International Conference on World Wide

Web Companion, pp. 83-84.

Reyaee, S. & Ahmed, A. (2015) Growth Pattern of Social

Media Usage in Arab Gulf States: An Analytical

Study. Social Networking, 4, pp. 23-32.

R- project. (2015) The R Project for Statistical Computing,

[Online] Available from: http://www.r-project.org/l.

[Accessed: 4th July 2015].

Shoukry, A. & Rafea, A. (2012) Sentence Level Arabic

Sentiment Analysis. Proceedings of the International

Conference on Collaboration Technologies and

Systems, Denver, USA, pp. 546 - 550.

Sokolova, M. Japkowicz, N. and Szpakowicz, S. (2006)

Beyond Accuracy, F-Score and ROC: A Family of

Discriminant Measures For Performance Evaluation.

In Hutchison, D. Kanade, T. Kittler, J. Kleinberg, J.M.

Mattern, F. Mitchell, J.C. Naor, M. Pandu Rangan, C.

Steffen, B. Terzopoulos, D. Tygar, D. & Weikum, G.

(Eds.). AI 2006: Advances in Artificial Intelligence.

Lecture Notes in Computer Science (4304). Australia:

Springer Berlin Heidelberg.

Stop Words (2014)

Stop-Words[Online] Available From:

Https://Code.Google.Com/P/Stop-Words/. [Accessed:

17th June 2015].

Storck, M. (2011) The Role of Social Media in Political

Mobilisation: a Case Study of the January 2011

Hybrid Sentiment Analyser for Arabic Tweets using R

423

Egyptian Uprising. [Online] Available from:

http://www.culturaldiplomacy.org/academy/content/pd

f/participant-papers/2012-02-bifef/The_Role_of_

Social_Media_in_Political_Mobilisation_-

_Madeline_Storck.pdf/. [Accessed: 2th July 2015].

Tsatsoulis, C. & Hofmann, M. (2014) Focusing On

Maximum Entropy Classification of Lyrics byTom

Waits. Proceedings of the IEEE InternationalAdvance

Computing Conference, Gurgaon, pp. 664 - 667.

Vinodhini G. & Chandrasekaran, RM. (2012) Sentiment

Analysis and Opinion Mining: A Survey. International

Journal of Advanced Research in Computer Science

and Software Engineering. (2). P. 283- 292. [Online]

Available from: http://www.dmi.unict.it/~faro/tesi/

sentiment_analysis/SA2.pdf. [Accessed: 16th June

2015].

KDIR 2015 - 7th International Conference on Knowledge Discovery and Information Retrieval

424