Social Data Sentiment Analysis in Smart Environments

Extending Dual Polarities for Crowd Pulse Capturing

Athena Vakali

, Despoina Chatzakou

, Vassiliki Koutsonikola

and Georgios Andreadis

Informatics Department, Aristotle University, Thessaloniki, Greece

School of Engineering, Aristotle University, Thessaloniki, Greece

Keywords: Sentiment Analysis, Social Data Processing and Social Networking, Microblogging Data Analysis, Mobile

Applications.

Abstract: Social networks drive todays opinion and content diffusion. Humans interact in social media on the basis of

their emotional states and it is important to capture people emotional scales for a particular theme. Such

interactions are facilitated and become evident in smart environments characterized by mobile devices and

new smart city contexts. This work proposes a sentiment analysis approach which extends positive and

negative polarity in higher and wider emotional scales to offer new smart services over mobile devices. A

particular methodology and a generic framework is outlined along with indicative mobile applications

which employs microblogging data analysis for chosen topics, locations and time. These applications

capture crowd pulse as expressed in microblogging platforms and such an analysis is beneficial for various

communities such as policy makers, authorities and the public.

1 INTRODUCTION

Social networks have drastically increased online

communication and human interactions, since

millions of users share opinions on a variety of

topics. Such activities embed both objective and

subjective criteria and certainly human reactions

govern social media diffusion. Sentiment analysis in

social media has gained considerable ground lately

since it facilitates human behaviour, responsiveness

and reactions understanding. Positive, negative and

neutral opinions are now declared, triggered and

visualized in most of the current social media

applications. Expressing such opinions and senses is

important to the markets and to stakeholders since

they can accordingly suggest, and apply policies and

services according to the social crowd opinions.

This work addresses the challenge to go beyond

such typical dual (positive and negative) polarity

since humans are certainly acting via wider and

more complicated emotional processes. Detecting,

summarizing and visualizing emotions in a

technically sound manner is important for capturing

social pulse, particularly when certain topics,

locations and timing are critical for decisions and

recommendations. Here, a method for extracting

social media affective knowledge is introduced on

the basis of a wider spectrum of six basic emotions

identified as seminal ones in the psychology

discipline. The purpose of this work is capturing the

crowd pulse by understanding people’s emotions as

expressed in social media platforms. Such capturing

is employed implicitly via computational methods

which overcome limitations of the dual

positive/negative analysis. Emotions’ intensity, is

surely important in social media activities and here

also the emotional states of socially circulated

information is considered with emphasis on the

smart contextual environment (such as a smart city).

The proposed analysis can be applied to a variety

of social media applications and initially

microblogging data streams are considered since

people interact in such platforms in an emotionally

driven freely manner which involves brief

information fragments but in an attitude inherent and

opinionated manner. Moreover, in microblogging

platforms (such as Twitter) people express their

opinions for different events which are underway in

smart different places of the world, and at different

time period. It is evident that mining and analysis of

microblogging data is important and necessary to

recognize interesting trends and opinions for

different topics and mobile smart devices offer

opportunities for instant opinion expression and

175

Vakali A., Chatzakou D., Koutsonikola V. and Andreadis G..

Social Data Sentiment Analysis in Smart Environments - Extending Dual Polarities for Crowd Pulse Capturing.

DOI: 10.5220/0004478401750182

In Proceedings of the 2nd International Conference on Data Technologies and Applications (DATA-2013), pages 175-182

ISBN: 978-989-8565-67-9

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

sentiment indications.

In this context, the contribution of this work is

summarized in the following:

 overcome dual (positive and negative) social

networking data analysis with the use of a wider

emotional spectrum;

 propose a flexible method which takes into

consideration many linguistic parameters, such

as the intensifiers and the valence shifters

(negation words), in order to understand the

emotions that are being expressed in micro-

blogs;

 characterize data fragments (microblogging

posts) in terms of their semantics in relevance to

a defined topic, their emotional intensity and

valence, by the use of particular lexicons and

dictionaries;

 propose a generic 3-tier framework which can

support mobile applications development with

inclusion of the proposed methodology which is

emotion-driven.

In the next sections social networks (and

microblogging in particular) sentiment analysis

efforts are summarized with emphasis on the role of

emotion to social networks data analysis. In Section

3, the role of emotions is emphasized for applying

affective analysis in social (microblogging) data

streams. In Section 4 a particular generic

implementation framework is outlined and an

indicative smart environment oriented mobile

application is highlighted in Section 5. Finally

conclusions and future work are summarized in

Section 6.

2 MICROBLOGGING

SENTIMENT ANALYSIS

AND BEYOND

Social networks offer services where users can post

information at anytime and anywhere in various

ways. Especially micro-blogs gain more and more

recognition for their real time characteristics and

short format. Extracting and analyzing the content of

them, so as to capture the emotional trends of

publishers, is considered too important. The goal of

affective analysis is to capture the wisdom of the

crowds, as well as the social pulse and the trends,

since information is valuable for improving social

network services in a wide spectrum of applications.

Defining the term “emotion” is a thoroughly

discussed issue and it certainly covers several angles

of human behaviour, especially when acting at a

social network. Here, we follow the definition given

by (Scherer, 2001), where emotion is defined as “an

episode of interrelated, synchronized changes in the

states of all or most of the five organismic

subsystems in response to the evaluation of an

external or internal stimulus event as relevant to

major concerns of the organism” with the five

subsystems being:

 information processing subsystem: is responsible

for internal and external stimulus evaluation. The

system is seen as continuously scanning the

environment and internal feedback signals in

order to determine the significance of stimulus

events;

 support subsystem: is responsible for the internal

regulation of the organism, especially the

generation of energy resources for action;

 executive subsystem: is involved in planning,

decision making and the preparation of action;

 action subsystem: is related to the

communication of reaction and intention through

motor expression and the execution of skeletal

movement for purposeful action;

 monitoring subsystem: is conceptualized as a

control system that reflects the current state of all

other subsystems.

These subsystems coordinate human behaviour and

are critical in human reactions so it is important to

proceed to sentiment analysis under their inclusion

in the process of social network data analysis. It is

moreover true that the above subsystems are

interrelated and any change in one subsystem will

lead to corresponding changes in others. So, it seems

reasonably to view emotion as a process with

constantly changing subsystem states. Work carried

out here focuses on capturing the social networking

activities which are of relevance to the information

processing and executive subsystems. This is

employed by integrating affective analysis which

focuses on emotions tracking (relevant to decision

making) with sentiment analysis which focuses on

opinions tracking (relevant to stimulus evaluation).

Microblog posts are of short text but they embed

links to other pieces of information (such as URLs)

and also users linkage can be tracked on the basis of

their common activity. Up to now sentiment analysis

in social networks was carried out targeting mostly

the level of the human emotion information

processing subsystem. For example, this is verified

by the fact that earlier approaches were scanning

human phrases to reveal and estimate their stimulus

triggering ability and they didn’t extend this to the

DATA2013-2ndInternationalConferenceonDataManagementTechnologiesandApplications

176

executive part which involves far more processes

(such as emotion-aware analysis and planning).

Typically, in earlier relevant work document and

resource level sentiment analysis sets as the

objective to determine whether an entire document

or a resource is positive, negative or neutral. This is

aimed in order to identify users stimulus and opinion

state. For example, in (Turney, 2002) an

unsupervised learning algorithm is presented for

classifying reviews into positive and negative ones.

This is carried out by initially extracting phrases

with adjective and adverbs and then estimating

semantic orientation of extracted phrases via a sum

of scores of each phrase. Pang et al. In (Pang, Lee,

and Vaithyanathan 2002) study the problem of

classifying documents based on overall sentiment

(positive/negative) of a document whereas in

(Benamara et al., 2007) an approach focuses on

text’s orientation and a sentiment analysis technique

uses a linguistic analysis of adverbs. Moreover, in

(Godbole et al., 2005) a system that assigns scores

indicating positive or negative opinions from texts

relevant to news and blogs is developed.

In sentence level sentiment analysis usually there

are two basic steps: determine the

subjectivity/objectivity score of each sentence and

further classify and determine whether subjective

sentences are positive or negative. In (Yessenalina

and Cardie 2011) new methods are presented

through which the sentences are categorized based

on how positive, negative or neutral they are. In (Pak

and Paroubek 2010); (O’Connor et al., 2010)

sentiment classifiers are utilized to determine how

positive, negative and neutral the messages derived

from Twitter are. More specifically, in (Pak and

Paroubek, 2010) they build a classifier that is able to

determine positive, negative and neutral sentiments

of tweets. In (O’Connor et al., 2010) a system is

introduced to compare the explicit knowledge taken

from twitter with the polls’ data (use of polls data as

ground truth). In (Yessenalina and Cardie 2011) the

orientation of each sentence of a text corpus is

recognized, but also the intensity of a text in a five-

scale system (very negative, negative, neutral,

positive, and very positive) is determined.

According to the authors’ knowledge, few of

earlier work has followed the direction of

categorizing microblogging based on some specific

primary or basic emotions. In (Gill et al., 2008) and

(Tsagkalidou et al., 2011) emotional classification

sets eight primary emotions (“fear, anger, disgust,

sadness, acceptance, anticipation, joy, surprise”) and

they proceed to affective analysis towards creating

groups of users that share the same emotions on

specific topics in Twitter. In the same context

(Bollen et al., 2010) use a Profile of Mood States

(POMS) as a psychometric instrument, and focus is

placed on six basic emotions, namely the “tension,

depression, anger, vigor, fatigue, confusion”, to

perform affective analysis of tweets.

This work addresses the open problem of

determining specific emotion scales along with their

intensity and valence in a social networking activity

(such as in a microblog), since this is important for

emotionally driven human reaction and execution

decisions, especially in todays smart environments.

The contribution of the proposed work is that users

opinions and orientation are estimated at a fine

grained level which considers the particular

emotions of importance to the information

processing and execution reactions.

3 ROLE OF EMOTIONS IN

MICROBLOGGING DATA

ANALYSIS

For successfully applying sentiment and affective

analysis it is very important to carefully design the

processes of relevance to human behaviour on one

hand and to computational needs on the other. This

is highly required in todays smart environments

which involve multi devices of mobile nature

utilized in technologically advanced contexts (such

as in smart cities).

Figure 1: Role of emotion in social networks ecosystem.

As depicted in Figure 1, at the real social

networking ecosystem many parameters and

activities are involved. If microblogging is used as a

case study, it is obvious that people in such

applications (e.g. in Twitter) post and highlight

SocialDataSentimentAnalysisinSmartEnvironments-ExtendingDualPolaritiesforCrowdPulseCapturing

177

information on the basis of their interests. They do

so at a particular time, place and context. This

emerging and often bursty activity is then perceived

by other people who proceed to specific actions on

the basis of their emotions and opinions. When

employing such microblogging data analysis it is

therefore important to increase awareness in terms of

users emotional drive and its role in decision making

Microblogging data are of brief nature (e.g. tweets

are of maximum 140 characters), and they embed

free text shortcomings mainly due to the users

freestyle and informal writing (e.g. abbreviations,

shortcuts, symbols etc). Therefore, language

inconsistencies are raising challenges in

microblogging data collections generation and no

qualitative guarantees can be ensured for a proper

sentiment and affective analysis.

It is obvious that out of a microblogging dataset,

some part of content will be emotionally relevant

and the rest has no contribution in the sentiment

and/or affect analysis. Therefore, an emotional

characterization of the data is needed in order to

keep and work with the particular data which carry

emotional information and is of importance for users

decisions and actions. To proceed at an emotional

aware microblogging data processing and

understanding, which will show a level of stability,

specific principles are needed and highlighted here.

Figure 2 summarizes the principles required in

sentiment and affect analysis. These principles aim

at understanding the specific emotion relevant

subsystems (described in Section 2) of information

processing and execution which are critical in

humans reaching decisions and actions. The

principles suggested here embed both qualitative and

quantitative criteria in order to cover human

emotional and computational scales. At the

qualitative part, the six emotional scale proposed by

(Eckman et al., 1982) is followed since it is widely

used in the bibliography and it has shown accurate

emotional capturing in text collections. At the

quantitative part, specific measurement for these

emotions are used in order to identify emotional

strength and orientation.

The proposed spectrum of emotions is defined on

the basis of six distinct primary emotions which

form the emotional states in a low-dimensional

space. The six primary emotions used are: “anger”,

“disgust”, “fear”, “joy”, “sadness”, and “surprise”

(Eckman et al., 1982). These emotions set the

ground for the microblogging data analysis, since by

using them each microblogging piece of data (e.g. a

preprocessed tweet content) can be comparatively

expressed with respect to each of these emotions.

Based on the above, we characterize the

emotional nature of each tweet by an emo(tweet)

function (Definition 2.1). This function is easily

used to calculate each tweet’s relevance to each of

the six primary emotions, taking into consideration

the final set of words which represent the tweet.

Figure 2: Sentiment and affect analysis principles.

Definition 2.1: The emotional proximity of a

particular tweet with respect to the particular

emotion scale of six emotions



6,...,1ie

defined by :

tweet words

emo (tweet) simscore( j, e )







(1)

The

),(

ejsimscore

value is proposed to capture

the “distance” of each word j of the tweet for each of

the





6,...,1



emotions. To estimate such

distance in practice any semantic lexicon (such as

Wordnet) can be utilized. Such lexicons are typically

organized in groups of words which are sets of

cognitive synonyms (so-called synsets) and on top of

them some similarity measures are suggested. These

similarities return a score which expresses the

semantic proximity of two words, i.e. two words are

considered as similar on the basis of the different

definitions of a word but also on the relationships

among word semantic taxonomies and hierarchies.

Two crucial parameters are used in order to

facilitate a computational procedure which will

complement semantics with emotional scaling:

 the intensity, which captures the degree of the

emotional excitement and here intensity is used to

define the tweet’s strength of the emotion, i.e. the

DATA2013-2ndInternationalConferenceonDataManagementTechnologiesandApplications

178

degree and power of a tweet expressed emotion. A

list of intensifiers is available in the bibliography

and a particular list with intensifier scores is used

here (Maite et al., 2010).

Definition 2.2: The emotional intensity of a

particular tweet is characterized by the tweet’s

words which are empowered intensifiers and it is

defined by:

tweet _ words

intens (tweet) int( j)*simscore(j, e )





(2)

where int(j)=1 only if the j word belongs to the

above intensifier set (else it is 0).

Equation (2) analyzes the intensity evaluation for

a tweet so it is obvious that only the words declared

as intensifiers contribute to the tweet’s intensity

characterization.

 the valence, which refers to the negative or

positive emotional value assigned by a person to

another person, event, goal, object and outcome,

based on its attractiveness. In our case valence is

used to captures the orientation of a tweet’s

emotion. This is employed by capturing

“semantic orientation” of positiveness or

negativeness embedded in a tweet’s word or

phrases. In practice this can be realized via the so

called valence shifters which are developed here

on the idea that typically, valence shifters reverse

word’s polarities (e.g. words like “not”, “aren’t”

etc) so their shifting capability should be

carefully considered.

Definition 2.3: The valence orientation of a

particular tweet is characterized by the tweet’s

words which are shifting and reverse polarities and it

is defined by:

tweet _ words

valens (tweet) val( j) * simscore( j, e )





(3)

where val(j)=1 only if the j word belongs to the

defined word shifters set (else it is 0).

The above defined measures enable extending

tweets semantics with capabilities of discovering

tweets emotional relevance as well as quantifying

emotional degrees. This is rather important since

social network users surely react driven by their

emotional excitement which leads to strong and

often shifting terms postings.

4 ROLE OF EMOTIONS IN

MICROBLOGGING DATA

ANALYSIS

Based on the introduced principles, a generic 3-tier

framework is outlined in Figure 3, involving

appropriate data collection, data analysis and

processing, as well as applications parts. Each of the

tiers interacts with its sequencing tier to proceed

from raw to emotionally relevant and clean data, and

from analysis and processing to particular

applications which can address specific criteria

(such as time, location and topic).

Figure 3: A 3-tier framework for emotion-aware

microblogging analysis and application.

4.1 The Data Collection Tier

The initial tier involves the collection of various

topic-driven datasets derived from microblogging

services (in our case Twitter) as depicted in Figure

3. The data collection is performed using the Twitter

streaming API which collects data on the basis of a

set of keywords which are representatives of the

topics used for analysis. The retrieved data includes

the actual tweets text as well as the timespan and

locality information that will be useful for further

analysis and knowledge extraction. Over such data

collections, the emotional evolution over time along

with the geographical distribution of emotions for a

particular topic can be examined.

The collected datasets can be relevant to various

topics in order to trigger and capture different

emotional behaviours in smart environments. The

proposed methodology is not restricted in a specific

topic set but it can be applied for every theme which

is characterized by a set of keywords. When data

collection is completed, an advanced processing

must take place that will lead to accurate and valid

results. This procedure involves the removal of the

semantically invalid information, eliminating, thus,

SocialDataSentimentAnalysisinSmartEnvironments-ExtendingDualPolaritiesforCrowdPulseCapturing

179

the existing noise. Without the application of an

extended processing of the datasets the resulting

methodology would be time consuming and less

stable. Thus, the datastreams should then proceed to

a particular pre-processing phase.

4.2 The Analysis and Processing Tier

Data collection process is followed by data

processing and analysis. The data calibration

constitutes a very important step in the whole

procedure, resulting to an appropriate data

formulation. As it was referred previously, the data

cleaning contains the removal of the tweet’s words

that are not semantically valid. Such semantically

invalid text involves words with no emotional

substance and also unusual words that do not

correspond to any English formulation according to

a dictionary (here we focus on an analysis outline

which uses content in the English language).

The next step is the similarity capturing of the

“clean tweets” in terms of the six primary emotions.

These primary emotions are able to capture the total

spectrum of expressed emotional situations of each

person in quite satisfactory extent. Given the

semantic lexicons, the calculation of the correlation

degree between the tweet’s words and the six

emotional states is based on the equation (1). For the

calculation of the emotional score of the tweet’s

words, emotional dictionaries are used. For the

scores computation, the intensifiers (equation 2) and

the valence shifters (equation 3) are taken into

consideration.

Given the calculated relations between tweets

and the six primary emotions the next step is the

data analysis through the data summary. Data

analysis may include various methodologies from

mining and machine learning. For example

algorithms such as k-means can be used for

grouping tweets with similar expressions towards

the six primary emotions. Here we follow a simple

classification approach that organizes the tweets by

defining a number of scales. Each scale represents a

different intensity level for each emotion for the

whole set of tweets. Organizing tweets in such a

manner is quite useful in recognizing patterns of

humans’ behaviour in relevance to different issues.

The proposed methodology is quite efficient in

capturing and understanding crowds’ emotions in an

implicit manner via computational methods. The

emotional aware clustering approach on the basis of

the wide spectrum of the primary emotions leads in

extraction of valid information that can be used later

for observing further conclusions.

4.3 The Implementation

and Applications Tier

At this tier particular visualization of the results can

be exposed in Web and/or mobile applications which

can range to various thematic disciplines and which

can be multi-criteria driven . More specifically such

criteria which can be taken into account in todays

social focused applications are location, time, and

topic (as emphasized in Figure 3). It is true that such

intelligent and collective information retrieval

methodologies can be used by a wide range of

applications which will integrate a geolocated

focused and time-aware system, based on a specific

topic. Another criterion is the operating platform, so

an application can be suitable for web, for mobiles

or for both of them.

The emotional patterns detected through the

proposed here methodology is of interest to a wide

range of market stakeholders, such as authorities,

companies, policy makers and the public. These

technologies can support companies to track the

current trends and adjust their communication

strategies, towards increasing quality of services and

their customers’ level of satisfaction. Additionally,

by applying such kind of methodologies in

microblogs’ data, the provided information is

organized in a more efficient way, making people’s

navigation and knowledge extraction easier and

more efficient.

5 CROWD PULSE DETECTION

IN SMART MOBILE

APPLICATIONS

Here, the proposed principles and methodology are

validated by mobile application tools which are

capturing the emotional patterns in microblogging

(twitter’s content) data streams over specific

contexts and smart locations. A more fine-grained

analysis is followed with widening emotions to the

six-class spectrum along with emotion intensity for a

particular topic/product, at a particular area and at a

particular time period. The proposed mobile

application offers further functionality with a map-

based representation, which displays distribution of

emotions for a particular topic and an appealing

visualization which depicts emotions’ dispersion of

a specific topic and/or a specific location. Two case

studies are summarized here, one for a smart city

orientation and the other for an entertainment topic.

These case studies are highlighted since they both

DATA2013-2ndInternationalConferenceonDataManagementTechnologiesandApplications

180

exhibit high emotional users involvement.

5.1 Smart City Pulses

The smart city case study is relevant to the area of

Santander in Spain which is implementing a large

smart city project (http://www.smartsantander.eu/).

Figure 4 summarizes the initial mobile screen of the

mobile application (implemented in Windows 7

environment), where the user can select between the

six primary emotions (anger, disgust, fear, joy,

sadness and surprise) at the Santander area.

Figure 4: A smart city emotion-aware mobile application.

Figure 5: A Smart City emotional dispersion for Surprise.

Upon user’s choice of a sentiment at the back

end of the application the proposed methodology is

applied and the sentiment analysis reveals the

dispersion of tweets in the particular smart city area.

The geographical visualization of tweets

facilitates emotion capturing and understanding. For

example, in Figure 5 the choice of the particular

emotion of “surprise” is depicted with particular

emotion dispersion which also summarizes the

quantitative microblogging activity with respect to

this emotion (expressed in number of twetts).

5.2 Crowd Pulse and Infotainment

Figure 6 summarizes the initial mobile screens of a

mobile application (available in Windows phone and

Android versions), where again user selects among

the six primary emotions (anger, disgust, fear, joy,

sadness and surprise) as well as the particular topic

(out of an indicative list). When the user has selected

a specific emotion, topic and location, a three

options navigation is allowed.

Figure 6: Emotion and criteria selection.

Figure 7 provides the first user’s option which

presents a graphical representation, in the form of

marker map that shows the distribution of emotions.

Each marker in the map represents the emotional

intensity of a tweet which appears on the map

according to its coordinates. The users can zoom-in

in order to see the coordinates and the intensity of

each tweet.

Another option for users will be the graphical

representation in the form of heat maps that

represent the intensity of emotions for a specific

emotion and topic in a particular area. Finally, the

user will have the possibility to see the distribution

of the emotions in city and country level in

relevance to a particular topic.

Figure 7: Location and emotion spectrum visualization.

The proposed application aims at covering

already available mobile tools which suffer from

qualitative results. This holds since the Web and

mobile market is dominated by the dual

SocialDataSentimentAnalysisinSmartEnvironments-ExtendingDualPolaritiesforCrowdPulseCapturing

181

(positive/negative) visualization of trends and

events, with only few of them embedding neutral

opinions. The innovation of the proposed tools is

that it is not limited in a positive-negative scale, but

it is extended in order to capture a wider spectrum of

humans’ emotions.

The graphical representation of humans’

emotions on maps leads to easier understandable and

efficiently organized results. The tool can be useful

for the identification of social trends and events’

impact. It can also provide an unprecedented level of

analytics for companies interested in promoting their

presence and products, authorities interested in

promoting a better way of living in particular

geographical context, and individual users

depending on their specific needs.

6 CONCLUSIONS

Micro-blogging services (especially Twitter) has

brought much attention recently as a hot research

topic in the domain of sentiment analysis. Existing

approaches mainly focus on the evaluation of tweets

emotional orientation on a dual basis i.e. positive or

negative. Our work, offers a 3-tier framework for

emotion-aware microblogging analysis, and extends

this emotional spectrum in six emotions, offering

thus a more fine-grained analysis of users’ emotions.

The overall process is based on emotional

dictionaries and considers linguistic parameters,

(intensifiers and valence shifters), to result in a more

accurate evaluation of the expressed emotions. The

proposed framework is the basis for mobile

applications which summarize and depict crowds’

emotions towards a specific topic and within a

certain locality. Such mobile application tools are of

great importance in capturing branding success,

diffusion in market and emotional states in relevance

to different topics (such as events, campaigns etc),

as expressed by people.

In the future we aim to extend our work by

incorporating more multi-language dictionaries that

will make possible the analysis of tweets written in

languages other than English and also to enhance

offered services to more areas and thematic

categories. Particular clustering algorithms are under

development for summarizing microblogging posts

in a more efficient manner.

REFERENCES

Benamara F., Cesarano C., and Reforgiato D., 2007.

“Sentiment Analysis: Adjectives and Adverbs are

better than Adjectives Alone”, AAAI International

Conference on Weblogs and Social Media Boulder,

CO USA.

Bollen J., Pepe A., and Mao H. 2010. “Modeling public

mood and emotion: Twitter sentiment and socio-

economic phenomena”, International Conference on

WWW2010, April 26-30, 2010, Raleigh, North

Carolina.

Ekman, P., Friesen, W. V., & Ellsworth, P., 1982. “What

emotion categories or dimensions can observers judge

from facial behavior?” In Emotion in the human face,

pp. 39-55. Cambridge University Press, 1982.

Gill A. J., French R. M., Gerle D. and Oberlander J., 2008.

“Identifying Emotional Characteristics from Short

Blog Texts”, Proc. of the 30th Annual Conference of

the Cognitive Science Society, 2237-2242, 2008.

Godbole N., Srinivasaiah M., and Skiena S., 2007. “Large-

Scale Sentiment Analysis for News and Blogs”, AAAI

International Conference on Weblogs and Social

Media (ICWSM’ 2007) Boulder, Colorado, USA.

Maite T., Julian B., Milan T., Kimblerly V., Manfred S.,

2010. “Lexicon-Based Methods for Sentiment

Analysis”, Association for Computational Linguistics.

O’ Connor B., Balasubramanyan R., Routledge B. R. and

Smith N. A., 2010. “From Tweets to Polls: Linking

Text Sentiment to Public Opinion Time Series”, Proc.

of the International AAAI conference on Weblogs and

Social Media, Washington DC, 2010

Pak A., Paroubek P., 2010. “Twitter as a corpus for

Sentiment Analysis and Opinion Mining”, Proc. of the

7th conference on International Language Resources

and Evaluation, 1320-1326, 2010.

Pang B., Lee L., and Vaithyanathan S., 2002. “Thumbs

up? Sentiment Classification using Machine Learning

Techniques”, Proc. of the Conference on Empirical

Methods in Natural Language Processing (EMNLP).

Scherer, K.R., 2001. “Appraisal Considered as a Process

of Multi-Level Sequential Checking”, in Appraisal

Processes in Emotion: Theory, Methods, Research, pp.

92–120, Oxford University Press.

Tsagkalidou K., Koutsonikola V., Vakali A., Kafetsios K.,

2011. “Emotional aware clustering on micro blogging

sources”, In ACII'11 Proceedings of the 4th

International conference on Affective Computing and

Intelligent Interaction pp.387-396.

Turney, P., 2002. “Thumbs Up or Thumbs Down?

Semantic Orientation Applied to Unsupervised

Classification of Reviews”, Proc. 40th Annual Meeting

of the Association for Computational Linguistics.

Yessenalina A., Cardie C., 2011. “Compositional Matrix-

Space Models for Sentiment Analysis”, Proc. of the

2011 Conference on Empirical Methods in Natural

Language Processing, pages 172-182, 2011.

DATA2013-2ndInternationalConferenceonDataManagementTechnologiesandApplications

182