ETL Transformation Algorithm for Facebook Opinion Data
Afef Walha
1
, Faiza Ghozzi
1,2
and Faïez Gargouri
1,2
1
Multimedia, InfoRmation Systems and Advanced Computing Laboratory, Sfax, Tunisia
2
Institute of Computer Science and Multimedia, Sfax University, Sfax, Tunisia
Keywords: ETL, Sentiment Analysis, Social Network.
Abstract: Considered as a rich source of information, social networking sites have been created lot of buzz because
people share and discuss their opinions freely. Sentiment analysis is used for knowing voice or response of
crowd for products, services, organizations, individuals, events, etc. Due to their importance, people
opinions are analyzed in several domains including information retrieval, semantic web, text mining. These
researches define new classification techniques to assign positive or negative opinion. Decisional systems
like WeBhouse, known by their data-consuming must be enriched by this kind of pertinent opinions to give
better help to decision makers. Nevertheless, cleaning and transformation processes recognized by several
approaches as a key of WeBhouse development, don’t deal with sentiment analysis. To fulfill this gap, we
propose a new analysis algorithm which determines user’s sentiment score of a post shared on the social
network Facebook. This algorithm analyzes user’s opinion depending on opinion terms and emoticons
included in his comments. This algorithm is integrated in transformation process of ETL approach.
1 INTRODUCTION
Sentiment analysis is concerned with the automatic
extraction of sentiment-related information from
text. Most sentiment analysis addresses commercial
tasks, such as extracting opinions from product
reviews. People can now post reviews of products at
merchant sites and express their views on almost
anything in social Websites.
With the growing availability and popularity of
opinion-rich resources such as social networking
sites (e.g. Facebook, Twitter), new opportunities and
challenges arise. In these sites, millions of users
interact frequently and share variety of digital
content with each other. They express their feelings
and opinions on every topic of interest. These
opinions carry import value for personal, academic
and commercial applications. Social networking
sites represent new and measurable sources of
information to an organization, such as customer’s
opinions on some products. These opinions may be
helpful for decision making.
Existing opinion analysis approaches propose
classification techniques and methods in order to
detect sentiment polarity. These approaches cover
many research domains including information
retrieval, text mining and semantic web. Decisional
systems can adopt proposed classification techniques
in order to integrate opinion analysis in Data
WeBhouse (DWB). Nevertheless, ETL design is
recognized as complex task. It is more and more
difficult including these techniques to analyze
pertinent opinion data.
In our previous works, we proposed an ETL
processes design approach integrating user’s opinion
available on Facebook social network. This
approach offers generic ETL operators to Webhouse
designer reducing the complexity of tackling opinion
extraction and transformation from Facebook
source. In this paper, we focus on opinion analysis
step of ETL transformation process which adapts a
lexicon sentiment analysis method. We propose an
algorithm that determines the user’s sentiment score
reflecting his opinion about a product or service
shared on Facebook pages. This score is resulted by
analyzing user’s comments based on lexical DB
composed of emoticons and opinion words
dictionaries.
This paper is organized as follow: section 2
presents a brief review on ETL design and opinion
analysis approaches. Then, we present an overview
of our ETL design approach integrating sentiment
analysis. In section 4, transformation process is
enriched by a new algorithm that combines visual
150
Walha A., Ghozzi F. and Gargouri F..
ETL Transformation Algorithm for Facebook Opinion Data.
DOI: 10.5220/0005494101500155
In Proceedings of the 11th International Conference on Web Information Systems and Technologies (WEBIST-2015), pages 150-155
ISBN: 978-989-758-106-9
Copyright
c
2015 SCITEPRESS (Science and Technology Publications, Lda.)
cues (emoticons) and opinion words collected from
user’s comments to determine his opinion polarity.
Finally, we conclude and present some perspectives
in section 5.
2 RELATED WORK
2.1 ETL Modelling Approaches
ETL processes design is a crucial task in DW
development due to its complexity and its time
consuming. Works dealing with this task can be
classified into two main groups: Specific ETL
modelling and Standard ETL modelling. The first
group offers specific notations and concepts to give
rise for new specialized modelling languages. ETL
processes proposed in (Vassiliadis, 2009) are limited
to typical activities (e.g. join, filter). (El-Sappagh et
al, 2011) extend these proposals by modelling
advanced operations, like user defined functions and
conversion into structure, etc. In order to design
complex ETL scenario, specific modelling
approaches propose conceptual or formal models.
However, the standardization is an essential asset in
modelling. The goal of the second group is to
overcome this problem by using standard languages
like UML, BPMN, etc. (Trujillo and Luján-Mora,
2003) and (Muñoz et al, 2010) use UML class
diagram to represent ETL processes statically or
dynamically by using UML activity diagram.
(Wilkinson et al, 2010) and (Akkaoui et al, 2012)
use BPMN standard where ETL processes can be a
particular type of business process.
Even though ETL modelling approaches
succeeded in providing interesting several modelling
methods and techniques, they don’t cover pertinent
opinion data sources available on web sources like
social networks, blogs, reviews, etc.
2.2 Opinion Analysis Approaches
Opinions are usually subjective expressions that
describe people sentiments, appraisals or feelings
toward entities, events and their properties.
Integrating opinion data is nowadays a hot topic
for many researchers. The common goal of
sentiment analysis approaches is to detect text
polarity: positive, negative or neutral. (Medhat et al,
2014) categorize opinion analysis approaches into
machine learning and lexicon approaches.
Machine learning approaches ((Wilson et al, 2005),
(Abbasi et al, 2008)) use classification techniques
(e.g. Naive Bayes, maximum entropy, and support
vector machines). Lexicon approaches rely on a
sentiment lexicon, a collection of known and
precompiled opinion terms. They use sentiment
dictionaries with opinion words and match them
with data to determine text polarity. They assign
sentiment scores to opinion words according to
positive or negative terms contained in the
dictionary. Lexicon approaches are divided into
dictionary-based approaches and corpus-based
approaches.
Dictionary-based approach ((Kim and Hovy,
2004), (Hu and Liu, 2004)) begins with a predefined
dictionary of positive and negative words, and then
uses word counts or other measures of word
incidence and frequency to score all opinions in the
data. The idea of these approaches is to first
manually collect a small set of opinion words with
known orientations (seed list), and then to grow this
set by searching in a known lexical DBs (e.g.
WordNet dictionary) for their synonyms and
antonyms. The newly found words are added to the
seed list (Liu, 2011). Opinion words share the same
orientation as their synonyms and opposite
orientations as their antonyms. (Qiu et al, 2010) and
(Hu and Li, 2011) use this technique to find
semantic orientation for adjectives. (Qiu et al, 2010)
worked on web forums to identify sentiment
sentences in contextual advertising.
Corpus based techniques rely on syntactic
patterns in large corpora. Corpus-based method can
produce opinion words with relatively high
accuracy. This method needs very large labeled
training data. (Jiao and Zhou, 2011) use Conditional
Random Fields methods in order to discriminate
sentiment polarity by multi-string pattern matching
algorithm applied on Chinese online reviews in
order to identify sentiment polarity. They established
emotional and opinion words dictionaries.
Machine learning and lexicon approaches use
opinion words and classification techniques to
determine text polarity. In addition to the use of
opinion words to analyze sentiment, emoticons
decorating a text can give a correct insight of the
sentence or text. For example, the emoticon “”
expressing “happiness” means positive opinion.
Further researchers take care of the increasing using
of these typographical symbols for sentiment
classification. In (Vashisht and Thakur, 2014),
authors identify the possible set of emoticons
majorly used by people on Facebook and use them
to classify the sentiment. Then, they use a finite state
machine to find out the polarity of the sentence or
paragraph. The problem with this approach is
performing sentiment analysis on text-based status
ETLTransformationAlgorithmforFacebookOpinionData
151
updates and comments, disregarding all verbal
information and using only emoticons to detect both
positive and negative opinions. (Hogenboom et al,
2013) propose a framework for automated sentiment
analysis, which takes into account information
conveyed by emoticons. The goal of this framework
is to detect emoticons, determine their sentiment,
and assign the associated sentiment to the affected
text in order to correctly classify the polarity of
natural language text as either positive or negative.
Existing ETL design approaches model various
web sources without considering user opinions
available on these sources including social networks,
reviews, blogs, forums or emails, etc. In the past few
years, many researchers have shown interest to
opinions expressed by people on any topic. They
proposed sentiment analysis methods and techniques
to determine text polarity. Some approaches apply
classification algorithms and use linguistic features
(machine learning approaches). Others use sentiment
dictionaries with opinion words and match them
with data sources to determine text polarity (lexicon
approaches). These approaches assign sentiment
scores to text according to positive or negative
words contained in the dictionary. Others
researchers use emoticons to disambiguate sentiment
when it is not conveyed by any clearly positive or
negative words in a text segment.
Sentiment analysis approaches presented in the
literature are very helpful and interesting to classify
text polarity. In spite of the importance of sentiment
classification approaches, we note that few of them
employ the coupling between sentiment analysis and
ETL processes in order to enhance semantic
orientation to multidimensional design. We propose
an ETL design approach adopting lexicon sentiment
analysis method. We consider Facebook opinion
data as a source to ETL processes. In the current
work, we define a new algorithm that analyzes
user’s comments about a product described on a
Facebook post and assign a sentiment score to him.
This score reflects user’s opinion. It is determined
based on emoticons and opinion words polarities
defined on lexical DB dictionaries.
3 ETL DESIGN APPROACH
OVERVIEW
In (Walha et al, 2015), we define a new ETL design
approach that integrates people’s opinions to model
Extraction, Transformation and Loading
processes. Figure 1 shows an overview of this
approach.
Extraction process starts by collecting general
Figure 1: Overview of ETL design approach.
WEBIST2015-11thInternationalConferenceonWebInformationSystemsandTechnologies
152
information about a post shared on a Facebook page.
A post is an individual entry of a user, page, or
group about a product or service. A list of available
actions (likes and comments) can be associated to
each post. These actions can help to gather people
opinions a post. The next step consists in identifying
users’ comments associated with the post.
Transformation process is organized in three
main steps: pre-processing, analysis and mapping.
Pre-processing involves comments cleaning and
comments’ opinion words and emoticons detection.
Analysis is the main step of transformation process.
It consists on analyzing user’s comments realized on
the post in order to classify his opinion (positive or
negative) about the product described in this post.
We propose for that an algorithm (section 4.1) which
assigns to each post a user’s sentiment score. This
proposal analyzes verbal cues (opinion words) and
visual cues (emoticons) contained in the post’s
comments. This algorithm adopts a lexicon
sentiment analysis method. It is based on emoticons
and opinion dictionaries defined in the lexical DB
(Walha et al, 2015). In these dictionaries, we
associate for each opinion term and emoticon a
sentiment polarity score which can be positive
(between (0) and (1)) or negative (between (-1) and
(0)). For examples, the polarity (0.9) is associated to
the opinion word “excellent” expressing
“Happiness” sentiment and the polarity (-0.7)
corresponds to the emoticon “” that expresses the
“sadness”. The final step in transformation process
is the mapping. It aims to match between ETL
source (concepts of “Facebook” model) and the
target (DWB model).
Loading process feeds the DWB with data
resulted from transformation step. It consists in
loading data into DWB multidimensional elements
including dimensions, measures, facts, attributes and
parameters.
4 TRANSFORMATION PROCESS
4.1 Opinion Analysis Algorithm
Transformation process analysis step (figure 1) aims
to detect users’ opinion according to their
comments. The goal of PostSentimentScore
(algorithm 1) is to determine a sentiment score
(SentP) to a post (P). SentP reflects opinion of the
user (U) about the product described in the post (P).
The user (U) can have a positive opinion if SentP is
comprised between (0) and (1), or negative opinion
with a value comprised between (-1) and (0). The
principle of algorithm 1 is to detect comments
realized by the user (U) on the post (P) and then
calculate their average, which corresponds to the
post sentiment score (SentP).
Algorithm 1: PostSentimentScore.
Input : P // A post shared on a Facebook page.
U // User who comments P.
Output : SentP // sentiment score assigned to the post P
1: SentP ← 0
2: N ← count (comments) // numbers of comments
published by the user (U) on the post (P)
3: For each Ci associated to the post P shared by U do
SentP ←SentP+CommentSentimentScore (Ci)
EndFor
3: return SentP / N
Algorithm 2: CommentSentimentScore.
Input : C // A comment shared by the user (U) on (P)
Output : SentC // sentiment score of the comment (C)
1: w ← countOpinionWords (C) // number of opinion
words in C
2: e ← countEmoticons(C) )// number of emoticons in C
3: SentC ← 0
4: For each wj of the comment C do
SentWj ← getOpinionWordPolarity (wj) // get the
polarity of opinion word wj defined in opinion dictionary
If modifier (wj) = true then
SentMj ← getModifierPolarity (mj) // get the
polarity of the modifier mj defined in opinion dictionary
If SentMj > 0 then
S ← 1
Else
S ← -1
endIf
Else
SentMj ← 0
EndIf
SentC ← SentC + S * (|SentMj| + SentWj) / 2
EndFor
5: For each ej of the comment C do
SentEj ← getEmoticonPolarity (ej) // get the
polarity of emoticon ej defined in emoticon dictionary
SentC ← SentC + SentEj
EndFor
6: If w + e > 0 then
SentCi ←
SentCi / (w + e)
endIf
7: return SentC
The score of the comment (C) is determined by
algorithm 2, untitled CommentSentimentScore. Its
principle is the following. First, it computes the
numbers of emoticons (e) and opinion words (w)
contained in (C). Then, it initializes SentC, i.e.
sentiment score of the comment (C), to the value (0).
This score is increased by polarity scores of all
emoticons and opinion words used in (C). These
scores are defined in the lexical DB (emoticon and
ETLTransformationAlgorithmforFacebookOpinionData
153
opinion dictionaries). Comment opinion words can
be related to a modifier, which can change its
sentiment polarity (e.g. the modifier “not” in the
comment “not good” change the user’s opinion). For
that, we verify the existence of modifier (mj) related
to each opinion word (wi) used in (C). A modifier
(mj) may change the polarity of (wj). For that, we
define the variable (S). Its value depends on modifier
polarity score (SentMj). It is equal to (1) in case of
positive value of (SentMj). Otherwise, the value (-1)
is associated to (S). Comment’s sentiment score
(SentC) is added to the average of modifier’s
polarity absolute value (SentMj) and opinion word’s
polarity (SentWj) multiplied by (S).
In our approach, we combine the use of opinion
terms and emoticons to detect user’s opinion
expressed on a comment. (SentC) is then increased
by the sum of emoticons polarities and finally
divided by the sum of emoticons (e) and opinion
words (w) used in (C).
4.2 Transformation Prototype
Facebook data are collected through Facebook
Graph API Explorer tool (API, 2015). To integrate
user’s opinion in the ETL prototype, we use this tool
to extract information about a Facebook post,
including post name, message, created_time,
pageName, link, type, etc. Also, we obtain user’s
comments realized on a post. Data collected from
Facebook about posts and their users’ comments are
converted into XML files including USERS,
PAGES, PRODUCTS, and POSTS. These files
collection composes our XML source DB.
The main goal of our ETL transformation
process is to analyze user’s opinion through
(transformation sentiment analysis step). We
adopt a lexicon based opinion analysis method. We
propose for that a lexical DB composed of
emoticons and opinion dictionaries. These latter are
transformed into XML files containing opinion
word, modifier and emoticon, their associated
sentiment classes (e.g. happiness) and polarity
scores. Figure 2 (XML lexical DB) depicts three
lexical DB files “emoticonsSample.XML”,
“opinionWordsSample.XML” and
“modifiersSample.XML”.
Figure 2: Transformation prototype.
WEBIST2015-11thInternationalConferenceonWebInformationSystemsandTechnologies
154
In figure 2, a sample of post (P) is also presented in
“PostSample.XML” file. To detect users’ opinions
based on their comments on the post (P), we apply
PostSentimentScore (algorithm 1) which returns a
sentiment score for each user that comments (P).
This score depends on opinion words and emoticons
exploited in comments. For that, we apply
CommentSentimentScore (algorithm 2). Results
are depicted in figure 2 (Post Sentiment Scores).
5 CONCLUSION AND FUTURE
WORKS
Due to the importance of people’s opinions
expressed on social networks for decisional systems,
we worked on integrating them in ETL processes
design. In this paper, we focus on ETL
transformation process. We propose a new
algorithm which analyzes user’s opinions expressed
through comments about a post shared on the social
network Facebook. Its goal is to detect both positive
and negative polarity. We associate for that a
sentiment score depending on comment’s opinion
terms and emoticons. In the proposed algorithm,
sentiment analysis adopts a lexicon method based on
opinion and emoticons dictionaries.
As future works, we intend to enrich our lexical
DB in order to adapt context-specific opinion
analysis. Also, we will extend our ETL processes
design approach by integrating more opinion web
sources including clickstreams, web sites, and others
social networks.
REFERENCES
Abbasi, A., Chen, H., Salem, A., 2008. Sentiment
Analysis in Multiple Languages: Feature Selection for
Opinion Classification in Web Forums. In ACM
Transactions on Information Systems Journal.
Akkaoui, Z., E., Mazón, J., Vaisman, A. A., Zimányi, E.,
2012. BPMN-Based Conceptual Modeling of ETL
Processes. In DAWAK’12, 14th
International
Conference on Data Warehousing and Knowledge
Discovery, pages 1-14, Springer.
API, 2015. API Graph Explorer Tool, “https://
developers.facebook.com/tools/explorer”.
El-Sappagh, S., H., Hendawi, A., M., Bastawissy, A., H.,
2011. A proposed model for data warehouse ETL
processes. In Journal of King Saud University -
Computer and Information Sciences, pages 91-104,
Elsevier.
Hogenboom, A., Bal, D., Frasincar, F., 2013. Exploiting
Emoticons in Sentiment Analysis. In SAC’13, 28th
Annual ACM Symposium on Applied
Computing, pages 703-710.
Hu, M., Liu, B., 2004. Mining and summarizing customer
reviews. In KDD’04, international conference on
Knowledge Discovery and Data Mining, pages 168-
177, ACM.
Hu, Y., Li, W., 2011. Document Sentiment Classification
by Exploring Description Model of Topical Terms. In
Computer Speech Language Journal, pages 386-403,
Elsevier.
Jiao, J., Zhou, Y., 2011. Sentiment Polarity Analysis based
Multi Dictionary. In ICPST’11, International
Conference on Physics Science and Technology,
Elsevier.
Kim, S., Hovy, E., 2004. Determining the Sentiment of
Opinions. In COLING’04, 20th International
conference on Computational Linguistics.
Liu, B., 2011. Web Data Mining: Exploring Hyperlinks,
Contents, and Usage Data. Springer-Verlag Berlin
Heidelberg, 2
nd
Edition.
Medhat, W., Hassan, A., Korashy, H., 2014. Sentiment
Analysis Algorithms and Applications: A Survey. In
Ain Shams Engineering Journal, pages 1093-1113.
Muñoz, L., Mazón, J.N., Trujillo, J., 2010. A Family of
Experiments to Validate Measures for UML Activity
Diagrams of ETL Processes in Data Warehouse. In
Information & Software Technology, pages 1188-
1203, Elsevier.
Qiu, G., He, X., Zhang. F., Shi, Y., Bu, J., Chen, C., 2010.
DASA: Dissatisfaction-Oriented Advertising Based on
Sentiment Analysis. In Expert Systems with
Application Journal, pages 6182–6191, Elsevier.
Trujillo, J., Luján-Mora, S., 2003. A UML Based
Approach For Modeling ETL Processes in Data
Warehouses. In ER’03, 22nd International Conference
on Conceptual Modeling, pages 307-320, Springer.
Vashisht, S., Thakur, S., 2014. Facebook as a Corpus for
Emoticons-Based Sentiment Analysis. In IJETAE’14,
International Journal of Emerging Technology and
Advanced Engineering, pages 904-908.
Vassiliadis, P., 2009. A Survey of Extract–Transform–
Load Technology. In IJDWM’09, International
Journal of Data Warehousing & Mining, pages 1-27.
Walha, A., Ghozzi, F., Gargouri, F., 2015. ETL design
toward social network opinion analysis. In SERA’15,
13th IEEE/ACIS on Software Engineering, Reasearch,
Management and applications, Springer (to appear).
Wilkinson, K., Simitsis, A., Dayal, U., Castellanos, M.,
2010. Leveraging Business Process Models for ETL
Design. In ER’10, 29th International Conference on
Conceptual Modeling, Springer.
Wilson, T., Wiebe, J., Hoffmann, P., 2005. Recognizing
Contextual Polarity in Phrase-Level Sentiment
Analysis. In HLT'05, 2005 Conference on Human
Language Technology and Empirical Methods in
Natural Language Processing. Association for
Computational Linguistics, pages 347-354.
ETLTransformationAlgorithmforFacebookOpinionData
155
