See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/312559872

Techniques for sentiment analysis of Twitter data: A comprehensive survey

Conference Paper · April 2016

DOI: 10.1109/CCAA.2016.7813707

CITATIONS READS

7 500

2 authors:

Mitali Desai Mayuri A Mehta

Sardar Vallabhbhai National Institute of Technology Sarvajanik College of Engineering and Technology
3 PUBLICATIONS   8 CITATIONS    21 PUBLICATIONS   42 CITATIONS   

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Mitali Desai on 23 March 2018.

The user has requested enhancement of the downloaded file.

 International Conference on Computing, Communication and Automation (ICCCA2016)

Techniques for Sentiment Analysis of Twitter Data:

A Comprehensive Survey

Mitali Desai Mayuri A. Mehta

Computer Engineering Department Computer Engineering Department
Sarvajanik College of Engineering and Technology Sarvajanik College of Engineering and Technology
Surat, India Surat, India
mitalidesai17@gmail.com mayuri.mehta@scet.ac.in

Abstract— The World Wide Web has intensely evolved a subject, sentiment itself i.e. belief and object i.e. the topic
novel way for people to express their views and opinions about about which the subject has shared the sentiment. An object is
different topics, trends and issues. The user-generated content an entity that represents a definite person, item, product, issue,
present on different mediums such as internet forums, discussion event, topic or any organization [3-7]. Sentiment analysis is
groups, and blogs serves a concrete and substantial base for carried out at different levels ranging from coarse level to fine
decision making in various fields such as advertising, political level. The coarse level sentiment analysis determines the
polls, scientific surveys, market prediction and business sentiment of the whole manuscript or document. The fine level
intelligence. Sentiment analysis relates to the problem of mining sentiment analysis, whereas focuses on the attributes.
the sentiments from online available data and categorizing the
Sentiment analysis of Twitter data is carried out on sentence
opinion expressed by an author towards a particular entity into
at most three preset categories: positive, negative and neutral. In
level which comes in between coarse level and fine level. In
this paper, firstly we present the sentiment analysis process to the sentiment analysis process, the sentiments present in the
classify highly unstructured data on Twitter. Secondly, we text are of two types: Direct and Comparative. The direct
discuss various techniques to carryout sentiment analysis on sentiments in text are independent from other objects in the
Twitter data in detail. Moreover, we present the parametric same sentence [7]. For example “the picture quality of this
comparison of the discussed techniques based on our identified camera is great.” However, the comparative sentiments in the
parameters. text denote the comparison of different objects within the
same sentence. For example “car x is cheaper than car y.”
Keywords— Sentiment analysis; machine learning; opinion
mining; Twitter The existing sentiment analysis techniques are useful in
various applications such as disaster relief and humanitarian
I. INTRODUCTION assistance, marketing and trade predictions, checking political
polls, advertising market, scientific surveys, checking
Social Computing is an innovative and growing computing
customer loyalty, finding job opportunities, population health
exemplar for the analysis and modeling of social activities
care and understanding students’ learning experiences [1-7].
taking place on various platforms. It is used to produce
intellectual and interactive applications to derive efficient In this paper, we present a sentiment analysis process for
results [1]. The wide availability of social media sites provides Twitter data. Twitter is a micro-blogging site that is rapidly
individuals to share their sentiments or opinions about a growing in terms of number of users [8-9]. Moreover, Tweets
particular event, product or issue. Mining of such informal and are mostly public and limited to 140 characters that simplify
homogeneous data is highly useful to draw conclusions in the identification of emotions in text [9-12]. Though, the
various fields. Though, the highly unstructured format of the abundance of data, use of short forms, timing of different
opinion data available on web makes the mining process posts, and diversity of language make the sentiment analysis
challenging [2]. process difficult for Twitter data.
Textual information present on web is majorly classified The rest of the paper is organized as follows: In section II,
into either of the two categories: fact data and sentiment data we discuss the existing work in the field of sentiment analysis.
[3]. Fact data are the objective terminologies concerning Section III describes the methodology to carryout sentiment
different entities, issues or events. Whereas sentiment data are analysis. Section IV presents numerous supervised machine
the subjective terms, that define individual’s opinions or learning algorithms used to conduct sentiment analysis and
beliefs for a particular entity, product or event. Sentiment their comparison based on the identified parameters. Finally,
analysis is the process of recognizing and classifying different Section V specifies the conclusion and future directions.
sentiments conveyed online by the individuals to derive the
writer's approach towards a specific product, topic or event is II. RELATED WORK
positive, negative or neutral. Sentiment analysis has three In current years, a voluminous amount of research has
major component of study as follows: sentiment holder i.e. been conducted in the sentiment analysis domain. In [7],

ISBN: 978-1-5090-1666-2/16/$31.00 ©2016 IEEE 149

 International Conference on Computing, Communication and Automation (ICCCA2016)

authors have proposed a technique to classify students’ data provided as an input to the built classifier to classify the
generated on Twitter into various categories to encounter remaining data i.e. test set. Each of the processing steps is
students’ various problems. In [13], authors have presented the discussed thoroughly in the following sub-sections.
logical approach to mine the sentiments shared on different
social media platforms. They have analysed the sentiments of
the text using combinatory categorical grammar, annotation,
lexicon acquisition and semantic networks. The basic
techniques of sentiment classification and the methods for data
collection are presented in [14]. The accuracy of classification
process with selected feature vector is verified for the
electronic products domain using various classifiers such as
Nave Bayes, Maximum Entropy, Support Vector Machine,
and Ensemble classifiers in [15]. In [16], authors have
introduced a hybrid method that is a combination of the usage
of sentiment lexicons with a machine learning classifier for Fig. 1. Sentiment analysis process of Twitter data
polarity detection of subjective texts in the consumer-products A. Data Sources
domain. In [17], authors have proposed a batch of machine
learning methods with semantic analysis to classify the Selection of data source to conduct the sentiment analysis
sentence and reviews of different products based on twitter plays a significant role. Social media platforms as the data
data using WordNet for better accuracy. In [18], authors have sources are broadly categorized into three general categories:
examined the performance of different classifiers such as blogs, micro-blogging sites, and review site [13-16]. Among
Naïve Bayes, SMO, SVM and Random Forest to classify all categories, a micro-blogging site such as Twitter has
Twitter data. In [19], authors have presented a technique to gained higher popularity due to its limited strength of the
normalize the noisy or irrelevant tweets and classify them content and publically availability of data. From the following
according to the polarity i.e. positive or negative. Moreover, statistics of the Twitter growth rate, it’s evident to use Twitter
they have employed a mixture model approach to generate as the data source for sentiment analysis.
different sentimental words. The generated words were later  Twitter Growth Rate Statistics
used as feature indicators in the classification model. Authors
have introduced a novel method to predict sentiments about Approximately 6,000 tweets are tweeted on Twitter on per
stocks using various monetary communication boards and second basis. It resembles to 350,000 tweets sent per
performed an automatic prediction for the stock market using minute and 500 million tweets per day. That makes it
web sentiments in [20]. In [21], authors have examined the around 200 billion tweets per year. In Twitter's history, the
performance of sentiment analysis in e-learning domain using number of Tweets increased from 5,000 tweets per day in
various methods of feature selection i.e. CHI statistics, Mutual 2007 [8] to 500,000,000 tweets per day in 2013, that is
Information (MI) and Information Gain (IG). In [22], authors approximately a six orders of magnitude [8]. At the
have proposed an automatic sentiment classifier to classify intermediate stages it has the statistics of 300,000 tweets
reviews of Brazilian TV shows into positive or negative per day in 2008 [9], 2.5 million tweets per day in 2009 [9],
category and possessed 90% of accuracy. Authors have 35 million tweets per day in 2010 [8], 200 million tweets
demonstrated a system to extract the Tweets and classify them per day in 2011 [10]. And 340 million tweets per day six
using domain oriented seed based enrichment technique to years after the emergence of Twitter i.e. on March 21,
reduce the information loss in the knowledge domain in [23]. 2012 [12]. This statistics conclude the use of Twitter for
In [24], authors have investigated numerous combinations of our research.
different preprocessing levels, machine learning techniques
 Twitter Studies
and features combining with neutral class to analyze real-time
students’ feedback. In [25], authors have developed an As per the recent work, the studies carry out on Twitter
enhanced sentiment classification method that can detect and data are in the field of health care, marketing, politics,
remove anomalies from Twitter data in addition to the advertising market, athletics etc. Analysis techniques used
classification. in these studies include qualitative content analysis,
network or graph analysis, linguistic or psycholinguistic
III. METHODOLOGY FOR SENTIMENT ANALYSIS analysis, word clouds and histograms [5]. In addition,
The sentiment analysis of Twitter data is an emerging field Twitter has been voted as the most promising source for
that needs much more attention. Fig. 1 shows the steps to carry the studies such as community or influence detection, topic
out the process of sentiment analysis on Twitter data. discovery, market and business predictions,
recommendation systems and tweet classification.
Firstly, the collected Twitter data is pre-processed to
perform the data cleaning. Secondly, the important features  Tweets
are extracted from the clean text, applying any of the feature
The message posted on Twitter is called Tweet, which is
selection methods. Thirdly, the portion of the data is manually
limited to 140 characters. Tweets are generally composed
labeled as positive or negative Tweets to prepare a training set.
of one of the followings [10] [13] [14]: text, links,
Finally, the extracted features and the labeled training set are

150
 International Conference on Computing, Communication and Automation (ICCCA2016)

emoticons, and images. A six seconds video is even added  Terms Frequency and Term Presence: These features
as a Tweet component in 2012 [8-12]. Based on these denote individual and distinct words and their
components the mining is applied to classify text, links, occurrence counts.
images, emoji or emoticons and even videos. The Tweets
contains three notations including hashtags (#), retweets  Negative Phrases: The presence of negative words can
(RT) and account Id (@). change the meaning or orientation of the opinion. So it
is evident to take negative word orientation in account.
B. Twitter Data Collection Methods
 Parts Of Speech (POS): Finding nouns, verbs,
The three possible ways to collect Tweets for research are
adjectives etc. as they are significant gauges of
as follows [11]:
opinions.
 Data repositories such as UCI, Friendster, Kdnuggets,
and SNAP E. Sentiment Classification Techniques
There are typically two techniques to identify sentiment of
 APIs: Twitter provides two types of APIs such as the text [7] [13] [26-32]: knowledge based technique and
search API and stream API. Search API is used to machine learning techniques.
collect Twitter data on the basis of hashtags and stream
API is used to stream real time data from Twitter Knowledge based technique is also called Lexicon based
technique. The lexicon-based technique focuses on deriving
 Automated tools that are further classified into the opinion based lexicons from the text and then identifying
premium tools such as Radian6 [18], Sysmos, the polarity of those lexicons. Lexicons are the collection of
Simplify360, Lithium and non-premium tools such as known and precompiled sentiment terms. This approach is
Keyhole, Topsy, Tagboard and SocialMention further classified into Dictionary-based approach and Corpus-
C. Data Preprocessing based approach. In the Dictionary-based approach, we find the
opinion oriented words, and then examine the dictionary to
Mining of Twitter data is a challenging task. The collected collect their synonyms and antonyms. Whereas in the Corpus-
data is raw data. In order to apply classifier, it is essential to based approach, we create a list of opinion words and then
pre-process or clean the raw data. The pre-processing task based on their context specific orientations, we find additional
involves uniform casing, removal of hashtags and other related opinion words in a vast corpus. To conduct lexicon
Twitter notations (@, RT), emoticons, URLs, stop words, approach, a trivial set of words describing opinions is
decompression of slang words and compression of elongated collected manually with their known orientations as a mean of
words. The following steps show the pre-processing pre-processing task. The set is then grown gradually by
procedure. searching in the distinguished and widely used lexicon
 Remove the Twitter notations such as hashtags (#), dictionary tool such as WordNet or Sentiful for their
retweets (RT), and account Id (@). synonyms and antonyms [17-18].

 Remove the URLs, hyperlinks and emoticon. It is Whereas the main objective of machine learning
necessary to remove non letter data and symbols as we techniques is to develop the algorithm that optimizes the
are dealing with only text data. performance of the system using training data such as
examples and/or past knowledge and experiences. The
 Remove the stop words such as are, is, am etc. The machine learning provides a solution of the sentiment
stop words do not emphasize on any emotions, it is classification problem in two sequential steps:
intended to remove them to compress the dataset.
1) Develop and train the model using training set data i.e.
 Compress the elongated words such as happyyy into already labeled data.
happy.
2) Classifying the unlabeled or unclassified data based on
 Decompress the slag words such as g8, f9. Generally the trained or skilled model.
slang words are adjectives or nouns and they contain
the extreme level of sentiments. So it is necessary to Machine learning techniques are further classified into
decompress them. supervised and unsupervised techniques [13] [15] [26-30]. To
carry out sentiment analysis, typically the supervised machine
D. Feature Extraction learning techniques are used as we are dealing with subjective
The pre-processed dataset has various discrete properties. data. Supervised machine learning techniques highly depend
In feature extraction methods, we extract different aspects on training data which are already labeled data unlike in the
such as adjectives, verbs and nouns and later these aspects are case of unsupervised machine learning techniques. Based on
identified as positive or negative to detect the polarity of the the provided training data, the classifier will classify the rest
whole sentence. Followings are the widely used Feature data i.e. test data. A large number of supervised machine
Extraction methods. learning algorithms such as Logistic Regression, Naïve Bayes,
Decision Tree, Support Vector Machine (SVM), Random
Forest, Maximum Entropy, and Bayesian Network are used

151
 International Conference on Computing, Communication and Automation (ICCCA2016)

for sentiment analysis [7] [13-26]. Choice of an appropriate

algorithm for selected data and domain is a crucial step. (3)
IV. SUPERVISED MACHINE LEARNING ALGORITHMS FOR Where fi(c, d) is a feature, ƛi is a parameter to be predicted
SENTIMENT ANALYSIS and Z(d) is a normalization function. Unlike NB, maximum
From out in-depth study of the supervised machine entropy doesn’t make any assumption regarding the feature
learning algorithms, it has been observed that the following independency.
machine learning algorithms are widely used and give average
The motivating idea behind maximum entropy is building
accuracy in majority of domains as well as with different types
a uniform model that satisfies all the given constrains. For
of data. Moreover, they provide consistent average speed of
example, consider a four-way text classification problem with
classification process irrespective of the size of input data
a constraint given as: on average 40% of documents having a
while handling the outliers.
word “professor" is labeled as faculty class. Innately, when
A. Naïve Bayes (NB) Approach given a document with “professor” word within, we assume
Naïve Bayes classifier [7] [17] [21] [26-30] is a simple that it has a 40% chance to be labeled as a faculty class, and a
probabilistic classifier that uses the concept of mixture models 20% chance to be labeled as each of the other three classes. If
to perform classification. The mixture model relies on the a document does not have “professor" word within then as per
assumption that each of the predefined classes is one of the the law of uniform class distribution, we assume the
components of the mixture itself. The components of the probability of that document to be in each class is 25%. This
mixture model denote the probability of belongingness of any model is precisely the maximum entropy model that conforms
term to the particular component. Thus, they are also known to each given or known constraint [30].
as generative classifiers. Naïve Bayes classifier is a C. Support Vector Machine (SVM)
probabilistic classifier that uses the concept of Bayes Theorem
Support vector machine (SVM) solves the traditional text
and finds maximum prospect of probability of any word fitting
categorization problem effectively; generally outperforming
to a particular given or predefined class. The probability P is
Naïve Bayes as it supports the concept of maximum margin.
defined as follows:
The main principle of SVMs is to determine a linear separator
(1) that separates different classes in the search space with
maximum distance i.e. with maximum margin [13-17]. If we
represent the tweet using t, the hyper plane using h, and
Where Xi is a given term and c is a predefined class label. classes using a set Cj € {l, -1} into which the tweet has to be
During the training phase, the incidence counts of the words classified, the solution is written as follows equivalent to the
are collected and stored in the hash tables. NB approach sentiment of the tweet.
suffers from an assumption that the features are independent in (4)
the feature space.
As per the definition of probability, the document d is The idea of SVM is to determine a boundary or boundaries
classified into class c using following equation: that separate distinct clusters or groups of data. SVM performs
this task constructing a set of points and separating those
(2) points using mathematical formulas. Fig. 2 illustrates the data
flow of SVM.
B. Maximum Entropy
The maximum entropy relies on probability distribution
estimation technique to perform classification. In this
technique, firstly the categorized feature sets are converted
into definite vectors using any of the encoding schemes.
Secondly, this encoded vector is used to compute weights for
each of the extracted features that can collectively support in
determining the most prospective label for a feature set. It is
used for various natural language processing tasks such as text
classification. It depends on the probabilistic approach like
Naïve Bayes [26-30]. The fundamental concept of maximum
entropy is that if much information regarding the data is not
known, the distribution should be extremely uniform. This
constraint eliminates the probability of non-uniform
distribution. The probability is derived from the categorized
training data and denoted as expected values of extracted
features as follows: Fig. 2. Support Vector Machine (SVM) workflow [30]

152
 International Conference on Computing, Communication and Automation (ICCCA2016)

D. Random Forest From the parametric comparison shown in Table I, we can

Random Forest classifier is a tree-based classifier. It conclude that Naïve Bayes algorithms are the simplest and
consists of numerous classification trees that can be used to easiest to understand and implement compare to Support
predict the class label for a given data point based on the Vector Machine and Maximum Entropy. However, it suffers
categorical dependent variable [19]. For a given data point, from lower accuracy due to its simple Bayesian probability
each tree votes for a particular class label and the class label assumption. Whereas Support Vector Machine provides the
gaining the maximum votes will be assigned to that data point. better accuracy but it doesn’t support the automatic learning of
The error rate of this classifier depends on the correlation or features. Maximum Entropy provides the moderated accuracy
association among any two trees in the forest in addition to the but supports the automatic learning of features. Random
strength of definite or individual tree in the forest. In order to Forest is based on decision tree method, which gives high
minimize the error rate, the trees should be strong and the accuracy with automatic feature learning.
degree of associativity should be as less as possible. Though, the implementation accuracy of all these
In the classifier tree, the internal nodes are represented as algorithms highly depends on the numerous factors such as
the features, the edges leaving a node are represented as tests domain chosen, data source, amount of data and preprocessing
on the feature’s weight, and the leaves are represented as class method applied on the data.
categories. It performs classification preliminary from the root F. Evaluation Parameters
node and moves incrementally downward until a leaf node is
detected. The document is then classified in the category that In common, the performance of sentiment classification
labels the leaf node. This algorithm is used in many techniques is estimated using four indicators as follows:
applications of speech and language processing. Accuracy, Precision, Recall and F1-score [13-32]. These
indicators are computed using the confusion matrix given in
E. Evaluation of Supervised Machine Learning Algorithms Table II:
From our in-depth study of the above supervised machine
TABLE II. THE CONFUSION MATRIX
learning algorithms used to perform sentiment analysis, we
have identified several parameters such as understanding # Predicted Positives Predicted Negatives
complexity, theoretical accuracy, theoretical training speed,
performance with small number of observations and type of Actual Positive Number of True Positive Number of False
Cases Cases (TP) Negative Cases (FN)
the classifier.
Understanding complexity refers to the technical Actual Negative Number of False Positive Number of True Negative
Cases Cases (FP) Cases (TN)
difficulties to understand the algorithm. Theoretical accuracy
is the theoretical measure of how accurately the algorithm can
classify the test set data according to the provided training These indicators are defined by the following equations:
data. Theoretical training speed refers to how fast the data can
be trained. Performance is related to the accuracy of the (5)
algorithm. In general, accurate algorithms have good
performance. Classifier refers to the type of classifier the (6)
algorithm belongs. There are different types of classifiers such
as linear classifiers, probabilistic classifiers, decision based
classifier [7] [13] [26-32]. (7)

TABLE I. PARAMETRIC COMPARISON OF THE SUPERVISED MACHINE (8)

LEARNING ALGORITHMS
Algorithm NB SVM Maximum Random
Entropy Forest Accuracy is defined as all true predicted cases against all
Understanding Very less High Moderate Moderate predicted cases. If we receive 100% accuracy, it denotes that
complexity the predicted cases are precisely the same as the actual cases.
Precision is defined as the true positive predicted cases against
Theoretical Low High Moderate High
accuracy
all positive predicted cases. Recall is defined as the true
positive predicted cases against all actual positive cases. F1 is
Theoretical High High Moderate Low a harmonic average of the precision and the recall.
Training Speed
V. CONCLUSION AND FUTURE WORK
Performance High Low Low Low
with small no.
In this paper, we have firstly presented the detailed
of Observations procedure to carryout sentiment analysis process to classify
highly unstructured data of Twitter into positive or negative
Classifier Probabilistic Linear Probabilistic Tree Based categories. Secondly, we have discussed various techniques to
carryout sentiment analysis on Twitter data including
knowledge based technique and machine learning techniques.

153
 International Conference on Computing, Communication and Automation (ICCCA2016)

Moreover, we presented the parametric comparison of the [14] S. Bhuta, A. Doshi, U. Doshi and M. Narvekar, “A review of techniques
discussed supervised machine learning techniques based on for sentiment analysis Of Twitter data”, Issues and Challenges in
Intelligent Computing Techniques (ICICT), 2014, pp. 583-591.
our identified parameters. It has been found that various
[15] M. S. Neethu and R. Rajasree, “Sentiment Analysis in Twitter using
techniques applied for sentiment analysis are domain specific Machine Learning Techniques”, in 4th Int. Conf. of Computing,
and language specific. Communications and Networking Technologies (ICCCNT), 2013, pp. 1-
5.
Hence, the future opportunities in the domain of sentiment
[16] S. Bahrainian and A. Dangel, “Sentiment Analysis using Sentiment
analysis include developing a technique to perform sentiment Features”, in Int. joint Conf. of Web Intelligence and Intelligent Agent
classification that can be applicable to any data regardless of Technologies, 2013, pp. 26-29.
domain. In addition, language diversity in social media data is [17] G. Gautam and D. Yadav, “Sentiment analysis of twitter data using
a key issue which is required to be eliminated in future. machine learning approaches and semantic analysis”, in 7th Int. Conf. on
Moreover, some of the more crucial challenges of Natural Contemporary Computing, 2014, pp. 437-442.
Language Processing (NLP) can also be used as further [18] B. Gokulakrishnan, P. Plavnathan, R. Thiruchittampalam, A. Perera and
developments in sentiment analysis, such as hidden or veiled N. Prasath, “Opinion Mining and Sentiment Analysis on aTwitter Data
Stream”, in Int. Conf. on Advances in ICT for Engineering Regions,
sentiment detection, satire detection, comparison or 2012, pp. 182-188.
association handling and emoticon detection.
[19] A. Celikyilmaz, D. Hakkani-Tur and Junlan Feng, “Probabilistic model-
based sentiment analysis of twitter messages”, IEEE Spoken Language
REFERENCES Technology Workshop (SLT), 2010, pp. 79-84.
[1] I. King, J. Li and K. T. Chan, “A Brief Survey of Computational [20] V. Sehgal and C. Song, “SOPS: Stock Prediction Using Web
Approaches in Social Computing”, in Proc. of Int. Joint Conf. on Neural Sentiment”, in 7th IEEE Int. Conf. on Data Mining Workshop, 2007, pp.
Network, 2009, pp. 2699-2706. 21-26.
[2] S. R. Barahate and V. M. Shelake, “A Survey and Future Vision of Data [21] Z. Kechaou, B. M. Ammar and A. M. Alimi, “Improving e-learning with
mining in Educational Field”, in Proc. 2nd Int. Conf. on Advanced sentiment analysis of users' opinions”, in Global Engineering Education
Computing & Communication Technology, 2012, pp. 96-100. Conference (EDUCON), 2011, pp. 1032-1038.
[3] Bing Liu, N. Indurkhya and F. J. Damerau, Handbook of Natural [22] A.C.E.S Lima. and L.N. de Castro, “Automatic sentiment analysis of
Language Processing, Second Edition, 2010, pp. 1-3860-68. Twitter messages”, in 4th Int. Conf. on Computational Aspects of Social
[4] M. Dredze , “How Social Media Will Change Public Health”, IEEE Networks (CASoN), 2012, pp. 52-57.
Intelligent Systems, 2012, pp. 1541-1672. [23] R. Batool, A. M. Khattak, J. Maqbool and S. Lee, “Precise tweet
[5] G. Siemens and P. Long, “Penetrating the fog: Analytics in learning and classification and sentiment analysis”, in 12th Int. Conf. on Computer
education”, Educause Review, 2011, vol. 46, no. 5, pp. 30-32. and Information Science (ICIS), 2013, pp. 461-466.
[6] C. Romero and S. Ventura, "Educational Data Mining: A Review of the [24] N. Altrabsheh, M. Cocea and S. Fallahkhair, “Sentiment analysis:
State of the Art," in Systems, Man, and Cybernetics, Part C: towards a tool for analysing real-time students feedback”, in 26th
Applications and Reviews, IEEE Transactions, 2010, vol. 40, no.6, pp. International Conference on Tools with Artificial Intelligence, 2014, pp.
601-618. 420-423.
[7] X. Chen, M. Vorvoreanu and K. Madhavan, “Mining Social Media Data [25] Z. WANG, V. J. Chuan TONG, X. XIN and H. C. CHIN, “Anomaly
to Understand Students’ Learning Experiences”, IEEE Transaction, Detection through Enhanced Sentiment Analysis on Social Media Data”,
2014, vol. 7, no. 3, pp. 246-259. in 6th International Conference on Cloud Computing Technology and
[8] Weil, Kevin (VP of Product for Revenue and former Big Data engineer, Science, 2014, pp. 918-922.
Twitter Inc.), "Measuring Tweets." Twitter Official Blog, February 22, [26] V. Singh and S. K. Dubey, “Opinion mining and analysis: A literature
2010. [Online]. Available: http://www.internetlivestats.com/twitter- review” , in 5th Int. Conf. on Confluence The Next Generation
statistics. [Accessed: 19-Oct-2015]. Information Technology Summit (Confluence), 2014, pp. 232-239.
[9] Krikorian, Raffi (VP, Platform Engineering, Twitter Inc.), "New Tweets [27] K. Khan, B. Baharudin, A. Khan and F. Malik, “Mining Opinion from
per second record, and how!" Twitter Official Blog. August 16, Text Documents: A Survey”, Digital Ecosystems and Technologies,
2013.[Online]. Available: https:// blog.twitter.com/ 2013/ new-tweets- 2009, pp. 217-222.
per- second-record-and-how. [Accessed: 19-Oct-2015]. [28] K. Ghag and K. Shah, “Comparative analysis of the techniques for
[10] Twitter Engineering, "200 million Tweets per day." Twitter Official Sentiment Analysis”, in Int. Conf. on Advances in Technology and
Blog. June 30, 2011. [Online]. Available: Engineering, 2013, pp. 1-7.
https://blog.twitter.com/2011/200-million-tweets-per-day. [Accessed: [29] W. Medhat, A. Hassan and H. Korashy, "Sentiment analysis algorithms
19-Oct-2015]. and applications: A survey”, Ain Shams Engineering Journal, vol. 5, no.
[11] “Three Cool and Inexpensive Tools to Track Twitter Hashtags”, June 4, 2014, pp. 1093-1113.
11, 2013. [Online]. Available http://dannybrown.me/2013/06/11/three- [30] J. Khairnar and M. Kinikar, “Machine Learning Algorithms for Opinion
cool-toolstwitterhashtags/ [Accessed: 19-Oct-2015]. Mining and Sentiment Classification”, in International Journal of
[12] "Twitter turns six." Twitter Official Blog. March 21, 2012. [Online]. Scientific and Research Publications, vol. 3, no. 6, June 2013.
Available: https://blog.twitter.com/2012/twitter-turns-six. [Accessed: [31] A. Sarlan, C. Nadam and S. Basri, “Twitter Sentiment Analysis”, in Int.
19-Oct-2015]. Conf. on Information Technology and Multimedia, 2014, pp. 213-216.
[13] N. Kasture and P. Bhilare, “An Approach for Sentiment analysis on [32] P. Saloun, M. Hruzik and I. Zelinka, “Sentiment Analysis – e-Bussines
social networking sites”, Computing Communication Control and and e-Learning Common Issue”, in 11th IEEE Int. Conf. on Emerging
Automation (ICCUBEA), 2015, pp. 390-395. eLearning Technologies and Applications, 2013, pp. 339-34.

154

View publication stats