An Interpretation of Sentiment Analysis for

Enrichment of Business Intelligence

Bharat Singh Nidhi Kushwaha Om Prakash Vyas
Indian Institute of Information Indian Institute of Information Indian Institute of Information
Technology Allahabad Technology Allahabad Technology Allahabad
India 211012 India 211012 India 211012
Email: [email protected] Email: [email protected] Email: [email protected]

Abstract—Sentiment analysis plays a very important role in Recently, sentiment analysis have being getting attention by
BI’s (Business Intelligence) applications which has been evident psychologist, in which psychologists come together with the
in the recent market activities. Towards sentiment analysis for old custom of emotion research. Opinion mining and sentiment
most of the popular websites like Amazon, Facebook, Twitter
necessitate the review of the customers which are used as a analysis are unable to be escaped to the sentimental knowledge
feedback. It’s play very important role for product review, that try to recognize human emotions [5].
Business intelligence as well as in decision making. The main However, the online data have several flaws that potentially
problem that arises to the point of view of users/customers is that, hinder the process of sentiment analysis. The first flaw is that,
it is practically in-feasible to read all those online reviews one by since people can freely post their own content, the quality of
one, because some of the products might have tens of thousand
reviews. In this paper, reviews are collected from the sources like their opinions cannot be guaranteed. For example, instead of
Amazon, Flipkart, and then used a method to combine both NLP sharing topic-related opinions, online spammers post spam on
(Natural Language Processing) and machine learning approach. forum which are irrelevant to the item. Among these some
Word sense disambiguation is also considered for this study. An spams are meaningless at all, while others have irrelevant
improvised lesk algorithms is used for removing noise in the data. opinions also known as fake opinions [6] [7] [8]. The second
Different types of data have different types of properties and
therefore are suited to different techniques correspondingly. This flaw is that, ground truth of such online data is not always
problem is closely related to the large scale nature of social available. A ground truth is more like a tag of a certain
networks and the necessity to perform aggregation operations, opinion, indicating whether the opinion is positive, negative,
which results in the form of Pie-Chart. Thus, we aggregate or neutral. Therefore, the branch of sentiment analysis, pycho-
millions of reviews into more user-friendly format. logical emotion and affect-sensitive systems research must be
Index Terms—Sentiment analysis, Machine learning, Opinion
Mining, Web Mining, Pie-Chart, Business Intelligence.
carried out simultaneously.

I. I NTRODUCTION
Although in the current era, the Web has transformed
itself from read-only to read-write. This evolution attracted
interested users working together and giving out through social
networks side (like facebook, twitter, linkedIn etc), blogs,
wikis, online communities, and other collaborative medias.
A screen shot of reviews given by users on amazon.com
for Nikon camera is depicted in Figure 1. Collaborative
and collective knowledge has spread throughout the Web,
especially in the fields which belong to each and every day
life’s. Regardless of significant improvement, opinion mining
and sentiment analysis are still not recognized as new interdis-
ciplinary branches by research community [1] [2]. Computer
scientists and engineers apply machine learning algorithms for
explicit classification from streaming data, voice, video, and
text [3] [4]. Often, opinion mining and sentiment analysis, Figure 1. A screen shot of reviews given by users for Nikon Camera.
both two term are used interchangeably in some applications,
but sentiment analysis and opinion mining basically focus There are hundreds of websites offering the same product
on emotion recognition and polarity detection, respectively. at different prices and also with different user feedbacks.
Sentiment analysis and opinion mining are emerging fields A common person usually decides what to buy, but from
due to above reasons [1] [5]. where to buy remains the sole source of dilemma. If in

978-1-5090-2597-8/16/$31.00 2016
c IEEE 18
some way we can tell the user which website offers both [10] into 3 classes: Web content, Web structure and Web usage
affordable price with awesome user feedbacks, then we have mining. In addition, some researchers have also classify Web
hit the jackpot!. Various e-commerce sites like junglee.com and mining into two different approaches. In both, the classes are
buyhatke.com do only comparison based on prices at different sink from three to two: Web content and Web usage mining
sellers, but they completely ignore user reviews about their as depend on the applications, sometimes Web structure is
buying experience. We would explore this interesting problem considered as better part of Web Content [10]; secondly, Web
and offer consumers general opinions based suggestions. In Usage is considered as better part of Web Structure. As the
this paper, we focus on the reviews given by users to a core of data mining, all these three classes attract the attention
particular items, products, or gadgets. in the process of hidden patterns discovery of unidentified
Due to the unexpected growth rate of e-business websites important information from the Web data. All three mainly
as well as online users, the growth in the reviews tends to centered on different mining objects, based in contents of the
millions by shoppers. Therefore, an automatic review mining Web. Figure 2 depicts the Web classes and its objects.
which can handle these reviews, has turn into a prerequisite for In this manner, we present a concise introduction of each of
recent research topic. We aim to explore several data mining the classes. Knowledge discovery has been done through Web
techniques and implement them to mine the data of various content mining, in which the data are the tradition collection of
social networking websites. In the recent years, more and text documents, transactional records, microarray data and, in
more social networking platforms have been established. Some the recent, due to increasing growth of video, images, audios
prominent examples are Facebook, Flipkart, Amazon, Twitter, sites and usage which are implanted in or connected to the Web
Google+, and LinkedIn. page i.e. multimedia data. From the study of various related
These platform providers collect a huge amount of data topics, researchers have separated Web content mining with
for each of their users. The importance of opinion mining is two points of view: the database approach or/and the agent-
mainly because reviews belong to the different category and based approach as discussed in article [11].
contain different types of information which is very useful for
the Business Intelligence. The main consideration of this work
is to collect information from different resources with the help
of web crawler and remove redundant, noise and ambiguous
data. For this purpose data pre-processing, data cleaning and
pruning play a very important role and provide summarized
result in the form of Pie-Chart. The visual interpretation of
the thousands lines of reviews by thousand of reviewers play
very important role into the business intelligence and decision
making.
In the paper of Kosala et al. [9] they had recommended a
breakdown of Web Mining in the subsequent jobs of :
Figure 3. Task of the Opinion Mining
• Resource finding: the task of filtering projected Web data
sets, Information collection and pre-processing,
In section I, we had provided an introduction of the back-
• Generalization: explicitly determines general patterns at
ground of web mining classification [12]. In section II,
particular Web sites as well as across various sites and
we depicted sentiment analysis and background related to it.
• Analysis: validation or explanation of the mined patterns.
In this part, we have reviewed several paper that we have
considered in our work, and also mentioned the limitations of
some of them. The proposed method was described in III rd
section. In Section IV, we described experimental work, in
which we have given description of dataset as well as how the
method was implementation and what tools were used. We
also described the graphical results of our work. The paper is
concluded with conclusion in section V.

II. S ENTIMENT A NALYSIS AND I T ’ S BACKGROUND

Sentiment analysis or opinion mining [1] [2] [13], [14]
[15] is the process which refers to the application of natural
language processing (NLP), computational theory. Opinion
mining can be divided into three major categories w.r.t. lin-
Figure 2. Categorization of Web mining [10] guistic i.e. resources which refers the appraisal theory of any
document, second is the sentiment classification which refers
In general, Han and Kamber has categorised Web mining how can a review take place for any kind of review analysis,

2016 IEEE Region 10 Conference (TENCON) — Proceedings of the International Conference 19

and last one is opinion summarization. Task of opinion sum-
marization has been developed as a kind of algorithms which Step4: Result in the form of a graph.
is used to remove redundant data and give suitable and best
result for review mining. Step5: Repeat iteratively all steps.
Sentiment analysis have enriched literature, and world
widespread exploration on automatic text analysis for sen- With the help of this approach, taking decisions on individual
timent, for examples, sentiment classifiers [9] [16] [17] au- words (whether they are positive or negative in nature) and the
tomatic survey analysis [4] and opinion extraction [18]. aggregated words to provide decisions of which group belongs
All these methods extract the reactions type underlying into to a particular class. Thus, the overall result is more accurate
the data, it may be positive reaction, negative reaction or and non redundant.
somewhere in between i.e. neutral reaction. In Figure 3 task Sentiment classification is represented by the result in
of opinion mining has been described in brief. positive or negative meaning of a word and find a sense
Linguistic resource theory [2] is based on the previously for a particular word with these categories [9] [16] [17].
defined well set of data and appraisal theory [19] is based on Turney et al. [21] proposed a methodology in which they have
how a comment or review strongly done with the subjectivity given an approach to partitions the sentiment of users reviews
analysis. Purpose of this technique is to find a proper and using PMI. Here, PMI refers to Pointwise Mutual Information
suitable approach to determined reviews which could be a that is used in statistics and information theory to calculate
part of any sentiment classification. In this, the main purpose the association for text classification. PMI is a measure of
is to find how a user or developer put their thinking in the association between the attribute and a class. More specifically,
form of a review or comment [19]. This theory is also known this words are considered as attribute and its category is class.
as appraisal theory developed by J.R. Martin, a computer In Turney’s study, sentence often contains adverbs or adjec-
linguistic researcher. tives, so they collect all information and extract verb, adverb
Subjectivity analysis [20] of any review means to find out and adjectives with the help of a parser. If “x” represents the
which kind of data is suitable for consideration. In this, data words and “y” represents the category, then the formula for
is mainly present in the form of comment, so primary aim is PMI is given by
to find out which data is strongly commented for appraisal p(x, y)
theory, for example: “fine”, “good” and “best” is the example P M I(x, y) = log
p(x)p(y)
of positive comment or review and “bad”, “worst” are strongly
shows negative words. “Good” and “poor” are considered as a baseline terms for
Conjunction Method [5] evaluating the PMI. Sentiments are then categorized in their
respective manner. With the help of this method sentiment can
This method automatically expand linguistic resources for
be easily find out and accurate result can be achieved.
sentiment mining. This method consider a study of subjective
After applying certain steps on a training dataset the model
corpora that collect linguistic features which consider the
is capable to classify negative and positive polarity sentences.
appraisal theory of a sentiment feature [14] [15]. In this
The drawback of this system or methodology, it needs more
the authors consider that conjunction between two or more
time to process and to send queries to a web search engine.
adjectives gives an indirect information regarding reviews, this
In literature, we have found machine learning based senti-
study derived from a hypothesis that “conjunction usually has
ment analysis methods. These are document based polarity
same in nature, with the help of that they can easily find out
classification and topic-based categories with naive bayes and
similarity and difference in the given conjunction”.
maximum entropy, proposed by Pang et al [2] [22]. Yao Tian-
Their system find and utilized the indirect information with
fang et al. [3] had given DWC (Distance-Weight Count) to
the below steps:
check the polarity for an document. This paper is very useful
Step1: Find all conjunction of an adjective to find out a proper and actual sense for any comment in
that are taken or extract from the any website reviews. This approach also shows that which
given reviews and find relevant of comments is useful and important based on positive and
morphological relations. negative polarity. In the next section, we will discuss about
proposed approach.
Step2: Apply loglinear regression model
III. P ROPOSED A PPROACH FOR R EVIEW ’ S A NALYSIS
which combines needed information
from conjunction. Three main feature of our study are -
1) Collect reviews from other sources (websites: Ama-
Step3: After collecting all the relevant zon.com) via using Python code.
information from conjuction find 2) Use adapted LESK algorithm to remove word sense
out which word has the similar and disambiguation.
relative meaning in a particular 3) Using java meta programming and coreNLP to generate
sense. the results in the form of a Pie-chart.

20 2016 IEEE Region 10 Conference (TENCON) — Proceedings of the International Conference

 Objective data: Product’s reviews basically contain likes the examples of the polysemous word,
which is in the form of likes/dislike which can be easily count storing their contexts as lines in
and provide the explicit information for review analysis. an initially untagged data set.
Subjective data which is in the form of user comments, on
the basis of these comments one can try to find out which Step2: For each possible sense we are
data is positive or negative for the product. Sometime data collecting the small number of rat
is showing ambiguous behavior so we are trying to remove -ing examples in order to find out
this problem using some other algorithms. It is a implicit the actual sense of the word. This
representation of user feedback. is possible by using hand tagging
When we collect data from different resources, data can be a subset of training sentence.
redundant, so, after cleaning data our second step is to find
out the important and useful data. As we know that in English Step3a: Tagging all training examples, we
language (or any other natural language) sometimes sentences must contain or collects repetitive
or words can have different meanings means i.e. there can be of each sense and then applying the
more than one meaning of the same word . tagging method for the proposed
For example: Bank can be a financial institute or a river approach.
shore. Sometimes two completely different words can be same
but their meaning are different. For example Can, can be used Step3b: Train the classification algorithm
as model verb: You can do it, or as container: He brought a on particular seeds and apply them
can of cold drink. on different seeds.
This is called Homonymy. The difference between polysemy
and homonymy is not always confusing. Word sense disam- Step4: After collecting all the informat
biguation (WSD) is used to select the appropriate senses and -ion apply the resulting classifier
meaning of a word or sentence in a given context. So, we on the training set. Tagging of
will have to find the correct and appropriate meaning of any which data is not tagged.
word for a natural language. Word sense disambiguation is the
open problem of computational theory which is used for the Step5: Repeat the all step iteratively.
purpose of finding the right sense of a given sentence or word.
This process requires two main steps
1) A dictionary to specify which word can be used for a
particular sense
2) The corpus of language data, to find out which word is
disambiguated
The framework of our proposed method is shown in Figure
4. In the first step, the collection of reviews is done by crawler
from many review sites (like amazon.com) as an inputs for the
visual analysis, and apply some of the pre-processing steps to
make it feasible for our approach, it include Part of Speech
(POS), removal of stop words. Subsequently, we search the
attributes and the sentiment words which are kept in vector
space model. In addition, to find the polarity of reviews,
reviews are divided in three or two classes. It may be positive,
negative or neutral in three classes and positive or negative in
two class review polarity (depends on the application type of
reviews). Negative reviews show the indication for bad and
defective product. In general, positive reviews supposed to
be more supportive. Through the application of PMI [21], a
sentiment association groups are assembled. In the last step, Figure 4. The proposed approach Flowchart.
we achieve the graphical summaries related to each website’s
product or items. With the help of this summary customer-
oriented and reliable assessment will be produced. IV. E XPERIMENTAL R ESULTS AND D ISCUSSION
In our proposed approach we are try to improve over lesk
algorithm which are given as follow- Researchers and organizations across many applications
exploit the huge and unique source of data for various rea-
Input: Reviews collected from Web.
sons. The major principle for the development of the quality
Step1: In the given input, identify all services provided and improvement of deliverables are the user

2016 IEEE Region 10 Conference (TENCON) — Proceedings of the International Conference 21

sentiments. Blogs, micro blogs and review sites provide a good
understanding of the reception level of products and services.
A. Data Source
1) Blogs: Blogosphere terminology represents all the blogs
that are being posted on the online websites. Depending upon
the blog, users usually posted their opinions on the particular
topics or matter. These blogs can be utilized in many ways
for future trend prediction and review generation for a cold
user. Utilizing this source of opinions is commonly known as
sentiment analysis [22].
2) Review Sites: These sites contains customer reviews
which can be processed and mined for analysis of their
behaviours. Mainly crawler and scripting language (python)
are used for fetching these reviews from the online sources
or website’s pages. These reviews are generally written in
natural language, that contains positive and negative words.
Moreover, the polarity of one review can be determined
Figure 5. Result showing the reviews in graphical way. It shows the positive,
by sentiment analysis process. CNETdownload.com, Ama- negative and neutral review’s percentage in pie-chart regarding their features
zon.com, Flipkart.com, and eBay.com are the good examples camera, battery and performance. The first image shows the over all review
for the product review sites [23]. In this paper, we are only of the NIKON D5300 item.
considering review sites data of the famous (Amazon.com).
Therefore, in the Figure 5 we have shown experimental results B. Analysis and Assessment
on NIKON camera’s reviews data which was fetched from
Also, Apple iphone-5S reviews are collected from the
Amazon.com.
amazon.com website ans applying the proposed steps using
3) Micro-blogging: Microbloging is a communication ser-
our algorithm for producing results in Pie-chart as shown in
vice in which any user can post and broadcast message to
Figure 6.
the other users on the similar services. In 2005, tumblelogs
was the first microbloging application. Subsequently, various
services like twitter, Jaiku, pownce and Tumbler in 2007 and
Plunk in 2008 came into the picture as a micro-blogging
servises. Only on twitter hundreds of millions tweets posted
every day. In addition to this, many social media tools such
as Facebook, MySpace, Xing, LinkedIn are also generating
billions of microblogs daily [15], [24].
We implemented the proposed application in java, using
Netbeans IDE8.0.2. For NLP and sentiment, we have used
stanford CoreNLP and Senti-WordNet, respectively. To collect
data from different sources for our experimental task, we
implemented Web Crawler in python. Major benefit of this
technology is that we can reuse the useful component of this
application into another application and it mainly contains
various types of visual and non visual components.
After collecting all the reviews about a product or item,
these words are rated on a scale of -1 to 1 for distinguishing
them as negative, neutral and positive features separately.
Senti WordNet help us a lot to achieved this. In this paper, Figure 6. Reviews of Apple iphone shown in a graphical way. It shows the
we are only considering review sites data. In Figure 5 we positive, negative and neutral review’s percentage in pie-chart regarding their
features camera, battery and performance. The first pie-chart shows the over
have shown results on NIKON camera’s reviews data which all review of the Apple iphone-5S.
was fetched from amazon.com using python code. It shows
Positive, Negative and Neutral percentages in the Pie-chart We present to the customers Pie-charts, giving various
for all reviews given to a product. It will help to easily options with a optimized rating based on our review process-
summarized all the reviews just by analyzing Pie-chart. This ing techniques. Thus, helping the consumer to decide more
Pie-chart representation totally independent from number of informatively. Hence, in our study we find out that sentiments
reviews. play very important role in business intelligence, we consider
all the possible features regarding to our study, other than

22 2016 IEEE Region 10 Conference (TENCON) — Proceedings of the International Conference

that we must find out the possible solution of other problem [8] A. Mukherjee, B. Liu, and N. Glance, “Spotting fake reviewer
like fake data or the data which is not actual sometime some groups in consumer reviews,” in Proceedings of the 21st International
Conference on World Wide Web, ser. WWW ’12. New York,
organizations put the negative comments on the other website NY, USA: ACM, 2012, pp. 191–200. [Online]. Available: http:
which creates problem in actual data analysis. Thus, this //doi.acm.org/10.1145/2187836.2187863
problem should be removed. [9] R. Kosala and H. Blockeel, “Web mining research: A survey,” SIGKDD
Explor. Newsl., vol. 2, no. 1, pp. 1–15, Jun. 2000. [Online]. Available:
http://doi.acm.org/10.1145/360402.360406
V. C ONCLUSION [10] J. Han, Data Mining: Concepts and Techniques. San Francisco, CA,
USA: Morgan Kaufmann Publishers Inc., 2005.
The ultimate objective of this study was to use the reviews [11] R. Cooley, B. Mobasher, and J. Srivastava, “Web mining: information
and pattern discovery on the world wide web,” in Tools with Artificial
and comments obtained from users about online product, items Intelligence, 1997. Proceedings., Ninth IEEE International Conference
and any other item that are being sold or for selling on e- on, Nov 1997, pp. 558–567.
commerce sites. This study has led us to devise the way so [12] W. Jicheng, H. Yuan, W. Gangshan, and Z. Fuyan, “Web mining:
knowledge discovery on the web,” in Systems, Man, and Cybernetics,
that, instead of going through each comment one by one, a 1999. IEEE SMC ’99 Conference Proceedings. 1999 IEEE International
buyer or customer can quickly get the feedback of all users in Conference on, vol. 2, 1999, pp. 137–141 vol.2.
graphical format. It will assist buyers to buy good quality of [13] E. Cambria and A. Hussain, “Sentic computing techniques, tools, and
applications 2nd edition,” 2012.
product. [14] D. Olsher, “Full spectrum opinion mining: integrating domain, syntactic
As the online e-commerce trend grow, the number of cus- and lexical knowledge,” in Data Mining Workshops (ICDMW), 2012
tomers as well as its number of reviews also grow very high, IEEE 12th International Conference on, Dec 2012, pp. 693–700.
[15] E. Pak and P. Paroubek, “Twitter as a corpus for sentiment analysis
in such cases the proposed method work effectively for a and opinion mining,” in In Proceedings of the Seventh Conference on
customer. The challenge in sentiment analysis of items and International Language Resources and Evaluation, 2010.
products is to provide a brief summary expressed by the [16] B. Pang, L. Lee, and S. Vaithyanathan, “Thumbs up?: Sentiment
classification using machine learning techniques,” in Proceedings of
abundant number of users. Each user used different phrases the ACL-02 Conference on Empirical Methods in Natural Language
and words in giving their reviews. Sentiment analysis can be Processing - Volume 10, ser. EMNLP ’02. Stroudsburg, PA, USA:
carried out on a set of reviews and discovered attribute terms Association for Computational Linguistics, 2002, pp. 79–86. [Online].
Available: http://dx.doi.org/10.3115/1118693.1118704
dig out from the reviews. In our study, we found that two [17] R. M. Tong. (2001, Sep.) An Operational System for Detecting
types of approaches have been utilized for sentiment classifi- and Tracking Opinions in On-line Discussion. [Online]. Available:
cation: machine learning and semantic analysis. In the machine www.daviddlewis.com/events/otc2001/talks.html
[18] S. Morinaga, K. Yamanishi, K. Tateishi, and T. Fukushima, “Mining
learning approaches, reviews are collecting from different web product reputations on the web,” in Proceedings of the Eighth ACM
sites and train the dataset according to a particular classifier. In SIGKDD International Conference on Knowledge Discovery and Data
this paper, we performed sentiment analysis through proposed Mining, ser. KDD ’02. New York, NY, USA: ACM, 2002, pp. 341–349.
[Online]. Available: http://doi.acm.org/10.1145/775047.775098
method, which provides a graphical interpretation to assist the [19] J. Martin, English Text: System and Structure. Amsterdam, USA: John
perspective of business analysis, and thus results in product Benjamins Publishing, 1992.
improvement. [20] A. Fujii and T. Ishikawa, “A system for summarizing and visualizing
arguments in subjective documents: Toward supporting decision
making,” in Proceedings of the Workshop on Sentiment and Subjectivity
R EFERENCES in Text, ser. SST ’06. Stroudsburg, PA, USA: Association for
Computational Linguistics, 2006, pp. 15–22. [Online]. Available:
[1] E. Cambria, B. Schuller, Y. Xia, and C. Havasi, “New avenues in opinion http://dl.acm.org/citation.cfm?id=1654641.1654644
mining and sentiment analysis,” IEEE Intelligent Systems, vol. 28, no. 2, [21] P. D. Turney, “Thumbs up or thumbs down?: Semantic orientation
pp. 15–21, 2013. applied to unsupervised classification of reviews,” in Proceedings
[2] B. Pang and L. Lee, “Opinion mining and sentiment analysis,” Found. of the 40th Annual Meeting on Association for Computational
Trends Inf. Retr., vol. 2, no. 1-2, pp. 1–135, Jan. 2008. [Online]. Linguistics, ser. ACL ’02. Stroudsburg, PA, USA: Association for
Available: http://dx.doi.org/10.1561/1500000011 Computational Linguistics, 2002, pp. 417–424. [Online]. Available:
[3] B. Liu and L. Zhang, “A survey of opinion mining and sentiment http://dx.doi.org/10.3115/1073083.1073153
analysis,” in Mining Text Data, C. C. Aggarwal and C. Zhai, [22] V. Singh, R. Adhikari, and D. Mahata, “A clustering and opinion
Eds. Springer US, 2012, pp. 415–463. [Online]. Available: http: mining approach to socio-political analysis of the blogosphere,” in
//dx.doi.org/10.1007/978-1-4614-3223-4 13 Computational Intelligence and Computing Research (ICCIC), 2010
[4] H. Li and K. Yamanishi, “Mining from open answers in questionnaire IEEE International Conference on, Dec 2010, pp. 1–4.
data,” in Proceedings of the Seventh ACM SIGKDD International [23] A. Kaur and V. Gupta, “A survey on sentiment analysis and
Conference on Knowledge Discovery and Data Mining, ser. KDD ’01. opinion mining techniques,” Journal of Emerging Technologies in
New York, NY, USA: ACM, 2001, pp. 443–449. [Online]. Available: Web Intelligence, vol. 5, no. 4, 2013. [Online]. Available: http://ojs.
http://doi.acm.org/10.1145/502512.502579 academypublisher.com/index.php/jetwi/article/view/jetwi0504367371
[5] R. Agarwal, T. Prabhakar, and S. Chakrabarty, “i know what you feel: [24] P. Korenek and M. imko, “Sentiment analysis on microblog utilizing
Analyzing the role of conjunctions in automatic sentiment analysis,” appraisal theory,” World Wide Web, vol. 17, no. 4, pp. 847–867, 2014.
in Advances in Natural Language Processing, ser. Lecture Notes [Online]. Available: http://dx.doi.org/10.1007/s11280-013-0247-z
in Computer Science, B. Nordstrm and A. Ranta, Eds. Springer
Berlin Heidelberg, 2008, vol. 5221, pp. 28–39. [Online]. Available:
http://dx.doi.org/10.1007/978-3-540-85287-2 4
[6] B. Liu, “The science of detecting fake reviews,” 2014 (accessed Decem-
ber 7, 2014).
[7] N. Jindal and B. Liu, “Opinion spam and analysis,” in Proceedings of
the 2008 International Conference on Web Search and Data Mining,
ser. WSDM ’08. New York, NY, USA: ACM, 2008, pp. 219–230.
[Online]. Available: http://doi.acm.org/10.1145/1341531.1341560

2016 IEEE Region 10 Conference (TENCON) — Proceedings of the International Conference 23