An Analysis of Online Twitter Sentiment Surrounding the European

Refugee Crisis

David Pope and Josephine Grifﬁth

College of Engineering and Informatics National University of Ireland Galway, Galway, Ireland

Keywords:

Social Media Analysis, Sentiment Analysis, Refugee Crisis.

Abstract:

Using existing natural language and sentiment analysis techniques, this study explores different dimensions

of mood states of tweet content relating to the refugee crisis in Europe. The study has two main goals. The

ﬁrst goal is to compare the mood states of negative emotion, positive emotion, anger and anxiety across two

populations (English and German speaking). The second goal is to discover if a link exists between signiﬁcant

real-world events relating to the refugee crisis and online sentiment on Twitter. Gaining an insight into this

comparison and relationship can help us ﬁrstly, to better understand how these events shape public attitudes

towards refugees and secondly, how online expressions of emotion are affected by signiﬁcant events.

1 INTRODUCTION

Due to the rapid growth of online social media over

the last decade, and the ease of obtaining social media

data, many computing techniques within Information

Retrieval and machine learning domains have been

applied to social media data. Sentiment analysis tech-

niques are particularly relevant due to the nature of

social media data which often contains a diverse set

of human opinions.

From a high-level perspective, the goal of the ap-

plication of sentiment analysis (also known as opinion

mining) to social media data is to analyse the opin-

ions of online social media users with respect to a

particular topic or subject matter. According to Tu-

masjan et al. (Tumasjan et al., 2010) there are now

many streams and applications of sentiment analy-

sis research, examples of which are in the areas of

product marketing, project management and politics.

There are two main approaches to perform sentiment

analysis: a lexical-based approach in which large dic-

tionaries of psychologically evaluated words, terms

and word stems are used to calculate a number differ-

ent mood states (Tumasjan et al., 2010); and a super-

vised machine learning approach using classiﬁcation

techniques such as Support Vector machines, Naive

Bayes or Maximum Entropy to learn, given training

data, the sentiment associated with some new unseen

data (Thelwall et al., 2011).

In this paper, existing sentiment analysis tech-

niques are applied to a twitter data collection relating

to the current refugee crisis in Europe. This refugee

crisis has been described by organisations such as the

United Nations Refugee Agency (UNHCR) and the

European Commission for Humanitarian aid and Civil

Protection (ECHO) as the planet’s worst refugee cri-

sis since the second World War

. Sparked by violent

and brutal civil war in the Middle East, millions of

refugees have ﬂed to Europe in seek of shelter and

protection, particularly from Syria (UNS, 2016). This

migration has been received with varying levels of

emotions in Europe and it is these “varying levels of

emotions” that are under study in this paper.

The Twitter dataset, gathered over 68 days from

November 2015 to early January 2016, comprises of

English-language and German-language tweets, each

language representing roughly half of the dataset. The

aim of the work is ﬁrstly to ascertain if there are any

noticeable differences in sentiment across the English

and German populations (as represented by the two

tweet collections). In addition, during the collection

of the data, two noticeable events occurred in Eu-

rope: the Paris attacks on 13th December 2015 and

the Cologne attacks on the 31st December 2015. The

sentiment expressed in both datasets from dates sur-

rounding these two events will also be compared. The

sentiment categories chosen for the comparison are:

negative emotion, anger, anxiety and positive emo-

tion. A lexical approach (using Linguistic Inquiry

and Word Count, LIWC (Pennebaker et al., 2015b))

Syria crisis: Echo factsheet (online) posted May 2016.

Pope, D. and Grifﬁth, J.

An Analysis of Online Twitter Sentiment Surrounding the European Refugee Crisis.

DOI: 10.5220/0006051902990306

In Proceedings of the 8th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2016) - Volume 1: KDIR, pages 299-306

ISBN: 978-989-758-203-5

299

is taken to ensure that the ﬁndings are psychologically

strong with respect to the sentiment categories under

analysis.

To date, there has been little analysis of online

sentiment surrounding the current refugee crisis in

Europe. One recent study completed by Coletto et

al. used a supervised machine learning approach to

explore positive and negative polarized Twitter sen-

timent relating to the refugee crisis in Europe across

time and space (Coletto et al., 2016). In contrast to

Coletto et al.’s research, this study not only uses a dif-

ferent lexical approach to perform the sentiment anal-

ysis of tweets, and uses tweets in two languages, it

also uses two additional mood (anger and anxiety).

Thus the contributions of this work are in the gath-

ering and sentiment analysis of both an English and

German dataset relating to the refugee crisis in Eu-

rope and comparing the sentiment in these datasets

across languages and events.

The paper outline is as follows: previous work

relating to social media sentiment analysis is out-

lined in Section 2. The methodology is presented

in Section 3 where the data gathering, data cleansing

and LIWC categorisation of sentiment is described.

Two sets of related results are presented: Section

4.1 presents results on general trends found across

the two populations focusing on four sentiment cate-

gories; Section 4.2 focuses on two events in both pop-

ulations while again considering the same four senti-

ment categories and also considering word maps of

top-occurring terms on the days around the events.

2 RELATED WORK

The area of sentiment analysis or “opinion mining”

has received increased interest in Computer Science

and other disciplines over the last number of years.

Sentiment analysis techniques have been used to solve

a variety of problems. As Caragea et al. discuss

(Caragea et al., 2014), using sentiment analysis tech-

niques as forecasting or predictive tools can help re-

searchers, business analysts, business leaders, disaster

response personal, economists, politicians, journalists

and many more to extract and categorize raw data

from online social media and transform it into action-

able knowledge. With this newly obtained and valu-

able knowledge then comes effective decision making

in the problem domain, whatever domain it may be.

Examples include predicting future stock market

prices by analysing the mood of Twitter users (Bollen

et al., 2011) and predicting German political elections

with greater accuracy than traditional opinion polls or

surveys (Tumasjan et al., 2010). Asur et al. (Asur and

Huberman, 2010) have also demonstrated the predic-

tive power of social media in their study of forecasting

box ofﬁce revenues. They were successful at forecast-

ing box ofﬁce revenues for upcoming movie releases

better than any leading market-based predictors.

Li et al. also demonstrated the power of sentiment

analysis in their study exploring the relationship be-

tween mood and changes in the weather (Li et al.,

2014). Results indicated that a relationship did exist

between online sentiment on Twitter and changes in

the weather, most notably during periods of increased

snow depth. The results found that an increase in

snow depth correlated with an increase in two dimen-

sions of the proﬁle of mood state scale, Depression-

Dejection and Anger-Hostility.

Data from other social media platforms, such as

Facebook, have also been used as a platform to per-

form sentiment analysis. For example Kramer carried

out a sentiment analysis study exploring the gross na-

tional happiness of Facebook users (Kramer, 2010).

Gilbert and Karahalios used the older social media

platform LiveJournal to explore the relationship be-

tween widespread worry ion LiveJournal and ﬂuctua-

tions in stock market prices (Gilbert and Karahalios,

2009).

There are generally two main approaches taken in

performing a sentiment analysis study and include:

• Lexical Approach which involves examining the

semantic orientation of words and phrases in a

piece of text in order to attempt to extract and cal-

culate sentiment polarity. The lexical approach

incorporates the use of a lexicon or a dictionary

of pre-deﬁned words, word stems or phrases that

have been semantically tagged into a number of

different sentiment categories. Each word and

phrase in the lexicon is tagged with a score to sig-

nify mood intensity. Words, parts of words and

word stems contained within the lexicon can also

be categorized into other more explicit dimen-

sions of mood states such as anxiety and anger.

Tumasjan et al. used this approach to construct

psychological proﬁles of election candidates in

the 2011 German Federal election using LIWC

(Tumasjan et al., 2010).

• Machine Learning approach where typically a su-

pervised machine learning approach is used for

sentiment analysis which requires a labelled set

of training data. Human evaluators are tasked

with labelling a random sample of data obtained

from the corpus into the desired sentiment cate-

gories, i.e., negative, positive and neutral senti-

ment. The data produced by the human evalua-

tors is used as training data for the chosen super-

vised classiﬁcation algorithm to produce a model.

KDIR 2016 - 8th International Conference on Knowledge Discovery and Information Retrieval

300

Algorithms such as Naive Bayes or Support Vec-

tor Machines are often used. The remaining data

from the corpus is passed to the model to produce

the sentiment results. Thelwall et al. (Thelwall

et al., 2011) and Asure et al. (Asur and Huber-

man, 2010) used this approach in their studies.

3 METHODOLOGY

This section provides an overview of the methodol-

ogy used to produce the results of both the sentiment

analysis and term frequency stages. Firstly, English

and German tweets were gathered from Twitter. All

tweets gathered on a particular day are treated as a

unit of analysis and a sentiment score is obtained for

the tweets for each day. The tweet corpus is therefore

analysed based on the creation/publish date and the

language. Following this, the tweet corpus is passed

through a number of data pre-processing and ﬁlter-

ing techniques to prepare the data for the sentiment

analysis stage. The tweet corpus is then passed to the

LIWC tool which uses its own large internal lexicons

to produce the sentiment results. Sentiment scores

produced by the LIWC tool are recorded and stored

per day for each day in the time range. Using the

LIWC results the mean and standard deviations are

calculated to establish the signiﬁcance of the results

obtained. Finally a term frequency (TF) process is

performed for the days around two signiﬁcant events

to ﬁnd the top terms, and potential topics, being dis-

cussed by Twitter users on those days.

3.1 Data Gathering and Data

Pre-processing

Using the public Twitter REST search API, over 1.6

million tweets were gathered between the 6th Novem-

ber 2015 and 13th January 2016. All tweets gathered

were added to a tweet corpus of which 902,139 were

English tweets and 702,852 were German tweets. The

basic keywords used are listed in Table 1 and are all

variants of words used to describe the refugee crisis in

English and the corresponding German translations.

These keywords, and variations of the keywords in-

cluding capitalisation and the use of the “#” charac-

ter, are passed as parameter arguments to the Twitter

search API for English and German tweets.

Tweets were organised according to the language

used and the day of creation. All tweets per day, in

each language, were treated as a unit for future pro-

cessing. To help reduce the bias in the sentiment

analysis results, tweet objects that contain identical

unique identiﬁers and retweeted tweet objects were

Table 1: List of English and German Search Terms used for

Twitter Search API.

English Keywords German Keywords

refugee ﬂ

uchtling

refugee crisis ﬂ

uchtlingskrise

migrant crisis migrationskrise

economic migrant wirtschaftsﬂ

uchtling

asylum seeker asylbewerber

removed from the corpus. As the sentiment approach

used is lexical, the repeating of words in retweets,

which are in the sentiment dictionaries, would most

likely give a much higher sentiment score thus skew-

ing the results.

To help reduce the inﬂuence of noisy data on re-

sults, all tweets were passed through a number of

custom ﬁltering and data transformation techniques

to standardise the terms and symbols used across all

tweets. Hyperlinks and the characters “@” and “#”

were removed. The character “&” was transformed

to “and” or “und” depending on the language and

“U.S.A” was transformed to the full name. This step

was taken in accordance with instructions outlined in

the LIWC documentation (Pennebaker et al., 2015b).

Following the data pre-processing stages, 53,355

completely unique tweets remained of which 28,866

tweets were ﬂagged as English and 24,469 tweets

were ﬂagged as German.

3.2 Sentiment Analysis using a Lexical

Approach

The Linguistic Inquiry and Word Count (LIWC) tool

is a natural language text processing (NLTP) tool

that is available for academic use (Pennebaker et al.,

2015b) and has been used in previous studies of so-

cial media data to perform sentiment analysis (Tu-

masjan et al., 2010). LIWC uses a lexical approach

to perform sentiment analysis and the large LIWC

internal dictionaries, also known as lexicons, have

been reﬁned and developed over a number of years

by the LIWC development team which includes psy-

chologists and computer scientists (Pennebaker et al.,

2015b). It is for this reason that a lexical approach,

and speciﬁcally LIWC, was chosen for this study

rather than a machine learning approach. With a ma-

chine learning approach, the results would be depen-

dent on the quality of the training data. The train-

ing data in our approach would be difﬁcult to obtain,

given that the data is in both English and German and

that human evaluators may not always be able to la-

bel consistently according to the sentiment categories

used in the study. Although a number of sentiment

categories exist in LIWC, four LIWC categories were

An Analysis of Online Twitter Sentiment Surrounding the European Refugee Crisis

301

selected for the sentiment analysis results presented in

this paper and are listed here along with the words and

word stems (represented by “*”) which are associated

with those sentiment categories:

• Positive Emotion: words such as love, nice,

sweet, fantastic, heal, decent, honest, hope, word

stems such as ecsta* (ecstatic), encourag* (en-

courage), magniﬁc* (magniﬁcent), and emoticons

associated with positive emotion such as :), (:.

• Negative Emotion: words and word stems such

as agony, destruct, pain, resent, ignorant, dis-

satisf* (dissatisfaction), outrag* (outrage), vul-

nerab* (vulnerable) and suffering, and emoticons

associated with negative emotions such as :( are

also included in the negative emotion category.

• Anger: words and word stems such as hate, kill,

brutal, hostil* (hostile), rude, sinister, rape, prej-

udic* (prejudice), beaten, aggressive.

• Anxiety: words and word stems such as wor-

ried, feared, nervous, worry, anxious, afraid, em-

barras* (embarrassed), paranoi* (paranoid), sus-

pico* (suspicious).

LIWC also contains a mean score for each of the

above listed categories for both English and German

languages. These means are available for different

data sources, one of which is Twitter for the English

language (Pennebaker et al., 2015b) and general data

for the German language. We use these means as a

baseline to compare the sentiment scores of each ex-

periment as outlined in the LIWC 2015 and 2007 doc-

umentation guides (Pennebaker et al., 2015a) (Pen-

nebaker et al., 2007). We will refer to these means as

“LIWC mean” in the results. In addition, we also cal-

culate our own mean score based only on our dataset.

We refer to this mean as “Calculated mean” in the re-

sults.

The tweet object creation date was used to or-

der all tweet objects resulting from the pre-processing

stage. For each language, and for each day, the tweet

text contained within each tweet object for that day

was extracted and stored in a single ﬁle. Each text ﬁle

representing each day and language was then passed

to the LIWC tool. A category score for each of the

LIWC categories was calculated per day, per lan-

guage. LIWC generates this score based on the total

number of words present within the tweets that match

words, word stems, emoticons and expressions cate-

gorized within the speciﬁed categories of the English

and German internal LIWC dictionaries.

In addition to using the LIWC scores for the four

emotion categories, the analysis of two signiﬁcant

events is supported by the generation of weighted

word maps. These were produced using the top-30

words (with stop words removed) found per day and

graphically represented using the wordclouds tool

4 RESULTS AND DISCUSSION

Section 4.1 presents the LIWC results focusing on the

comparison across sentiment categories and popula-

tions. Section 4.2 also considers the same LIWC re-

sults, in addition to word maps, but focuses on only

two days in particular: 13th September 2015 and 31st

December 2015.

For all sentiment category graphs showing LIWC

scores over the time range of the study, the X axis rep-

resents the day in the time range and the Y axis con-

tains the daily scores produced by LIWC for a partic-

ular emotion category. Each sentiment category graph

contains two means per language: the LIWC mean is

the average LIWC mean (independent of the refugee

crisis data) and is unique for each category as already

discussed. The “calculated” mean represents the cal-

culated average score using the results produced by

the LIWC tool for the refugee crisis data gathered in

this study. Using this “calculated” mean the standard

deviation was also calculated for each category to al-

low the signiﬁcance of the LIWC results for each day

to be discussed. As is evident in the sentiment cate-

gory graphs in the results section, the results ﬂuctuate,

over time, for each day, demonstrating, for example,

an increase or decrease in a sentiment category.

4.1 Sentiment Category Results: per

Population and per Day

This section will present and discuss the LIWC results

in detail comparing both English and German results

together, per sentiment category: Negative Emotion,

Anger, Anxiety and Positive Emotion. Recall that

each tweet contains at least one of the search terms

relating to the refugee crisis, as outlined in Section

3.1.

• Negative Emotion. The LIWC results for neg-

ative emotion are illustrated in Figure 1 for both

English and German. The calculated means for

the English and German sets of results are both

above the LIWC mean. This suggests that, over

the entire datasets, for each day and in each lan-

guage, tweets relating to the refugee crisis show

a higher negative emotion level compared to the

LIWC averages. It can be seen in Figure 1 that

there are very few days that are below the LIWC

negative emotion mean. Also notable is the lack

Free online wordcloud generator.

KDIR 2016 - 8th International Conference on Knowledge Discovery and Information Retrieval

302

Figure 1: English and German LIWC Negative Emotion.

Figure 2: English and German LIWC Positive Emotion Levels.

of increased negative emotion after the 31st De-

cember in the German tweet data compared to the

English tweet data. While there is an increase in

expressed negative emotion in both English and

German tweet datasets after the 13th November,

there is a difference in the signiﬁcance of this in-

crease. For the English tweet data, there is an in-

crease of 2.5 standard deviations above the mean

after the 13th November whereas for the German

tweet data there is a sharper increase to over 5

standard deviations above the calculated mean.

However, in both cases there is a sharp decrease

in negative emotion following the increases.

• Positive Emotion. The LIWC results for posi-

tive emotion are illustrated in Figure 2 for both

the English and German datasets. Both calculated

means in Figure 2 are much lower than the LIWC

means. In fact positive emotion only surpasses

the LIWC mean for one day for both languages.

Thus, this suggests that across both result sets, for

English and German tweets, there is low positive

emotion expressed by Twitter users. This pattern

does correlate to the high negative emotion lev-

els expressed in Figure 1 previously. The positive

emotion levels in Figure 2 for the German tweet

data illustrate a much more aggressive ﬂuctuation

in scores across the time range under analysis.

• Anger. A similar pattern to that seen for negative

emotion levels can be seen in Figure 3 where the

sentiment scores for anger are mostly higher than

the LIWC means. This is highlighted in Figure

3 where the English tweet anger levels only drop

below the LIWC mean for one of the 68 days. The

German tweet anger levels indicate that there were

also a smaller number of days below the LIWC

mean. Across both the English and German tweet

datasets, a similar ﬂuctuation pattern in anger lev-

els over the entire time range is visible.

• Anxiety. The calculated mean values for anx-

iety levels for the English and German tweet

datasets are represented in Figure 4 where it can

be seen that both calculated means are above the

LIWC means. This suggests that across the en-

tire dataset, anxiety levels are higher than the ex-

pected LIWC mean. This was also seen with neg-

ative emotion and with anger levels. It is clear

from Figure 4 that for the English tweet dataset,

anxiety levels drop below the LIWC anxiety mean

An Analysis of Online Twitter Sentiment Surrounding the European Refugee Crisis

303

Figure 3: English and German LIWC Anger.

Figure 4: English and German LIWC Anxiety.

for only 5 of the 68 days within the time range un-

der analysis. Comparing the English and German

graphs in Figure 4 it can be see that, in contrast to

German tweet anxiety levels, English tweet anx-

iety levels show a more aggressive ﬂuctuation in

scores. In comparison, for the German tweets, a

period of relevant “calmness” following a period

of sharp increases and decreases of anxiety lev-

els between the 13th and 26th November, can be

seen.

4.2 Sentiment Category Results per

Event: Paris and Cologne Attacks

This section focuses on two events in particular: the

days after the 13th November 2015 and the days af-

ter the 31st December 2015 (following the night of

the Cologne attacks in Germany). In order to bet-

ter establish if Twitter users are discussing the events

that took place in Paris and Cologne, weighted word

maps, produced by calculating the frequency of the

top-30 terms, are generated. These weighted word

maps are shown in Figures 5 and 6 and illustrate

that in both the English and German result sets, the

Paris attacks feature as a top term in tweets pub-

lished after the event on the 13th November 2015.

The top German terms include “refugee coordinator”

(ﬂ

uchtlingskoordinator ), “federal government” (bun-

desregierung) and “paris”.

References to the events that occurred in Cologne

do not begin to surface in the English top terms un-

til the 7th January, almost a week after the event oc-

curred. This is also the case for the German dataset

where references to the Cologne attacks do not sur-

face until the 2nd January and only begin to appear

in the top terms on the 4th January 2016. The list

of top German terms include “d

anemark” (denmark),

“deutschland” (germany), “cologne” (k

oln), “leads”

uhrt), “passport control” (passkontrollen), “de-

ported” (abgeschoben) and “shots” (sch

usse). This

is an interesting ﬁnding, as it was reported that there

was a delay in the reporting of the events in Cologne

(Scally, 2016).

With respect to the sentiment category of Negative

Emotion, perhaps what is most obvious in Figure 1 is

the increase in Negative Emotion after 13th Novem-

ber, the night of the Paris terrorist attacks. The En-

glish tweet results in Figure 1 show an increase of 2.5

standard deviations above the calculated mean and the

KDIR 2016 - 8th International Conference on Knowledge Discovery and Information Retrieval

304

Figure 5: Top English Terms on 14th November 2015 and Top German Terms on 15th November 2015.

Figure 6: Top English Terms on 7th January 2016 and Top German Terms on 4th January 2016.

German tweet results show negativity levels of 5 stan-

dard deviations above the calculated mean after the

13th November. However, already mentioned, a sharp

decrease in negative emotion quickly follows the in-

creases after 13th November for both the English and

German datasets. What is also interesting in Figure

1 is the difference in negative emotion after the 31st

December 2015. For the English tweets, we can see a

gradual increase in negative emotion (up to 2.5 stan-

dard deviations above the calculated mean), whereas

there is little change in negative emotion levels in the

German tweets.

With respect to the sentiment category of Anger,

Figure 3 clearly illustrates an increase in anger af-

ter the 13th November and after the 31st January for

both German and English results, where in both cases,

scores reach 3 standard deviations above the calcu-

lated means.

With respect to the sentiment category of anxi-

ety, as shown in Figure 4, there is a prominent in-

crease in anxiety levels in German tweets after the

13th November 2015 reaching a high of over 5 stan-

dard deviations above the calculated mean. In con-

trast to German anxiety levels, English Tweet anxi-

ety levels display a more aggressive ﬂuctuation with

an increase in anxiety levels after the 13th November

2015 in the English tweet dataset. These higher peaks

of anxiety in the English tweets appear to decrease

and increase in intensity over a number of days and

weeks after the 13th November, ﬁnally decreasing to

the LIWC mean and calculated mean after the 7th De-

cember. There is evidence of a period of heightened

anxiety in English tweets between the 13th November

and the 7th December 2015. It is also interesting to

note a level of “calmness” in the German tweet scores

after the 25th November where there is no obvious or

signiﬁcant increase or decrease in anxiety levels, even

after the Cologne attacks on 31st December 2015.

5 CONCLUSION AND FUTURE

WORK

Using existing sentiment analysis techniques, the goal

of this study was to explore different mood states of

tweet content relating to the refugee crisis in Europe.

In addition, the goal of identifying changes in online

expressions of emotion in tweet content triggered by

signiﬁcant ofﬂine events relating to the refugee crisis

An Analysis of Online Twitter Sentiment Surrounding the European Refugee Crisis

305

was also undertaken in this study.

The sentiment categories of negative emotion,

positive emotion, anger and anxiety were analysed

across two populations (English and German speak-

ing) and across 68 days. Two signiﬁcant events oc-

curred during these 68 days (the Paris Terrorist at-

tacks and the Cologne attacks) and these events were

analysed by considering the four sentiment categories

in addition to the frequency of words used in tweets

around those days. A lexical approach using the

LIWC tool was adopted which, in addition to produc-

ing sentiment scores per category and per language,

provided mean scores for each category.

The two main goals of this study were achieved.

Firstly the results from the sentiment analysis stage

show some interesting trends and commonalities

across languages and days. Coupling these sentiment

analysis results with the results of the term frequency

analysis the second goal of determining if online twit-

ter sentiment is affected by signiﬁcant events relating

to the refugee crisis was also achieved. The results of

this study show interesting trends and commonalities

across languages and days. However, it is important

that we do not base conclusions about refugees and

the refugee crisis in Europe on subjective opinions.

Instead, using existing sentiment analysis techniques

the results of this study may help us to better under-

stand how online expressions of emotion and attitudes

towards refugees are impacted by signiﬁcant ofﬂine

events such as the Paris and Cologne attacks.

Other LIWC categories will be explored in future

work. It is evident in many of the sentiment graph re-

sults that there are a number of other spikes of height-

ened expressed emotion. In order to establish the top-

ics of discussion on these days the researchers are cur-

rently looking into topic detection using unsupervised

machine learning clustering techniques.

REFERENCES

(2016). Syria regional refugee response inter-agency infor-

mation sharing portal. [Online Website: Accessed:

2016-06-15].

Asur, S. and Huberman, B. (2010). Predicting the future

with social media. Web Intelligence and Intelligent

Agent Technology (WI-IAT), 2010 IEEE/WIC/ACM

International Conference.

Bollen, J., Mao, H., and Zeng, X. (2011). Twitter mood

predicts the stock market. Journal of Computational

Science.

Caragea, C., Squicciarini, A., Stehle, S., Neppalli, K., and

Tapia, A. (2014). Mapping moods: Geo-mapped sen-

timent analysis during hurricane sandy. 11th Interna-

tional Conference on Information Systems for Crisis

Response and Management.

Coletto, M., Lucchese, C., Muntean, C., Nardini, F. M.,

Esuli, A., Renso, C., and Perego, R. (May 2016).

Sentiment-enhanced multidimensional analysis of on-

line social networks: Perception of the mediterranean

refugees crisis. Under Review.

Gilbert, E. and Karahalios, K. (2009). Widespread worry

and the stock market. 4th International AAAI Confer-

ence on Weblogs and Social Media.

Kramer, A. (2010). An unobtrusive behavioral model of

“gross national happiness”. Proc. of SIG CHI Confer-

ence on Human Factors in Computing Systems.

Li, J., Wang, X., and Hovy, E. (2014). What a nasty

day: Exploring mood-weather relationship from twit-

ter. Proc. of 23rd ACM International Conference on

Conference on Information and Knowledge Manage-

ment.

Pennebaker, J., Boyd, R., Jordan, K., and Blackburn, K.

(August 2015a). The development and psychometric

properties of liwc2015’.

Pennebaker, J., Chung, C., Ireland, M., Gonzales, A.,

Booth, R., and Francis, M. (2007). The development

and psychometric properties of liwc2007.

Pennebaker, J., Roger, J., Booth, R., Boyd, L., and Fran-

cis, M. (August 2015b). Linguistic inquiry and word

count: Liwc2015, operator manual.

Scally, D. (2016). Media under scrutiny over slow response

to cologne attacks. The Irish Times Newspaper.

Thelwall, M., Buckley, K., and Paltoglou, G. (2011). Senti-

ment in twitter events. Journal of the American Soci-

ety for Information Science and Technology, 62.

Tumasjan, A., Sprenger, T., Sandner, P., and Welpe, I. M.

(2010). Predicting elections with twitter: What 140

characters reveal about political sentiment. 4th Inter-

national AAAI Conference on Weblogs and Social Me-

dia.

KDIR 2016 - 8th International Conference on Knowledge Discovery and Information Retrieval

306