Anger is More Influential Than Joy: Sentiment Correlation in Weibo

Abstract

Recent years have witnessed the tremendous growth of the online social media.
In China, Weibo, a Twitter-like service, has attracted more than 500 million
users in less than four years. Connected by online social ties, different users
influence each other emotionally. We find the correlation of anger among users
is significantly higher than that of joy, which indicates that angry emotion
could spread more quickly and broadly in the network. While the correlation of
sadness is surprisingly low and highly fluctuated. Moreover, there is a
stronger sentiment correlation between a pair of users if they share more
interactions. And users with larger number of friends posses more significant
sentiment influence to their neighborhoods. Our findings could provide insights
for modeling sentiment influence and propagation in online social networks.

Full text

arXiv:1309.2402v1 [cs.SI] 10 Sep 2013
Anger is More Influential Than Joy: Sentiment
Correlatio n in Weibo
Rui Fan, Jichang Zhao , Yan Chen and Ke Xu
1
State Key Laboratory of Software Development Environment, Beihang University,
Beijing 100191, P.R.China
Abstract
Recent years have witnessed the tremendous growth of the online social me-
dia. In China, Weibo, a Twitter-like service, has attracted more than 500
million users in less than four years. Connected by online social ties, dif-
ferent users influence each other emotionally. We find the correlation of
anger among users is significantly higher than that of joy, which indicates
that angry emotion could spread mor e quickly and broadly in the network.
While the correlation of sadness is surprisingly low and highly fluctuated.
Moreover, there is a stronger sentiment correlation between a pair of users if
they share more interactions. And users with larger number of friends posses
more significant sentiment influence to their neighborhoo ds. Our findings
could provide insights for modeling sentiment influence and propagation in
online social networks.
Keywords:
Sent iment influence, Emotion propagation, Sentiment analysis, Online
social network, Weibo
1. Introduction
From the view of conventional social theory, hom ophily leads to connec-
tions in social network, as the saying “Birds of a feather flock t ogether”
states [1]. Even in online social network, more and more evidence indicates
that the users with similar properties would be connected in the future with
1
To whom the correspondence should be addressed: kexu@nlsde.buaa.edu.cn.
Preprint submitted to Elsevier September 11, 2013

high probabilities [2, 3]. It is clear that homophily could affect user behav-
ior both online and offline [4, 5], while the records in online social network
are relatively easier to be tracked and collected. Moreover, the continuous
growth of the online social media attracts a vast number of internet users
and produces many huge social networks. Twitter
2
, a microblogging website
launched in 2006 , has over 200 million active users, with over 140 million
microblog posts, known as tweets, being posted everyday. In China, Weibo
3
,
a Twitter-like service launched in 2009, has accumulated more than 500 mil-
lion registered users in less than four years. Everyday there will be more than
100 million Chinese tweets published. The high-dimension content generated
by millons of globa l users is a “big data ” window [6] to investigate the on-
line social networks. That is to say, these large-scale online social networks
provide an unprecedented opportunity for the study of human behavior.
Beyond typical demographic features such as age, race, hometown, com-
mon friends and interest, homophily also includes psychological states, like
loneliness and happiness [1, 7, 4]. Previous studies also show that the
computer-mediated emotional communication is similar to the traditional
face-to-face communication, which means there is no evident indicatio n that
human communication in online social media is less emotional or less per-
sonally [8]. The tweets posted in online social networks deliver not only the
factual information but also the sentiment of the users, which represents their
reflections on different social events. R ecent study [4] shows that happiness is
assortative in Twitter network and [6] finds that the average happiness scores
are positively correlated between the Twitter users connected by one, two or
three social ties. While in these studies, the human emotion is simplified
to two classes of positive and negative or just a score of general happiness,
neglecting the detailed aspects of human sentiment, especially the negative
emotion. Because of oversimplification of the emotion classification, it is
hard f or the previous literature to disclose the different correlat ions of differ-
ent sentiments and then make comparisons. However, the negative emotions,
like anger, sadness or disgust, are more applicable in real world scenarios such
as abnormal event detection or emergency spread tracking. Figuring out the
correlation of these emotions could shed light on why and how the a bno r ma l
event begins to spread in the network and then leads to large-scale collective
2
www.twitter.com
3
www.weibo.com
2

behavior acro ss the entire network. On the other hand, the investigation
of how the local structure affects the emotion correlation is not systemati-
cally performed yet, while which is essential to studying the mechanism of
sentiment influence and contagion.
Aiming at fill these vital gaps, we divide the sentiment of a person into
four categories, including ange r, joy, sadness and disgust, and investigate
the emotion correlation between connected users in the interaction network
obtained from Weibo. Out of our expectation, it is found that anger has a
stronger correlatio n between different users than that o f joy, while s adness’s
correlation is trivial. This indicates that ange r could propagates fast and
broadly in the network, which could explain why the real-world events about
food security, government bribery o r demolition scandal are always the hot
trend in internet of China. Moreover, node degree a nd tie strength both
could positively boost the emotion cor r elat ion in online social networks. Fi-
nally, We make our datasets in this paper public available to the research
community.
The r est of this paper is organized as follows. In Section 2, closely related
literature would be reviewed, including the methods of sentiment a nalysis
and the difference between our contributions and the previous findings. The
data sets employed in this paper and the methods of emotion classification
would be introduced in Section 3. We also define the correlation of emotion
in this section. Section 4 reports our findings in detail and our empirical
explanations and several real-wo r ld case studies would be elucidated in Sec-
tion 5. Finally, we give a further discussion in Section 6 and then conclude
this paper briefly.
2. Related works
The content in online social media like Twitter or Weibo is mainly recorded
in the form of text. Many approaches have been presented to mine sentiment
from these texts in recent years. One of them is the lexicon based method,
in which the sentiment of a tweet is determined by counting the number
of sentimental words, i.e., positive terms and negative terms. For exam-
ple, Dodds and Danforth measured the happiness of songs, blogs and presi-
dent s [9]. They also employed Amazon Mechanical Turk to score over 10000
unique English words on an integer scale from 1 to 9, where 1 represents
sad and 9 represents happiness [10]. Golder and Macy collected 509 million
English tweets from 2.4 million users in Twitter, then measured the posi-
3

tive and negative affects using Linguistic Inquiry and Word Count(LIWC)
(http://www.liwc.net). While another one is the machine learning based
solution, in which different features are considered to perform the task of
classification, including terms, smileys, emoticons and etc. The first step is
taken by Pang et al. in [11], they treat the sentiment classification of movie
reviews simply as a text categor izat ion problem and investigate several typi-
cal classification algorithms. According to the experimental results, machine
learning based classifiers outperform the baseline method based on human
words list [12, 13, 14]. Different from most work which j ust categorized the
emotion into negative and positive, [15] divided the sentiment into four
classes, then presented a framework based on emoticons without manually-
labelled training tweets and achieved a convincing precision. Because of
the ability of multi-emotions classification, we employ this framework in the
present paper.
Each user in the online so cial network could be a social sensor and the
huge amount of tweets convey complicated signals about the users and the
real-world events, a mo ng which the sentiments are an essential part. Emotio n
states of the users play a key role in understanding the user behaviors in so-
cial networks, whether from an individual or gro up perspective. In addition,
users’ mood states are significantly affected by the real-world events [16]. [17 ]
employed the public mood stat es to predict the stock market and [15] found
the variation of the emotion could be used to detect the abnormal event
in real-world, especially the negative sentiment. Individual happiness was
measured and several temporal patterns of happiness were revealed in [10].
In [18], Golder and Macy collected 509 million English tweets fro m 2.4 mil-
lion users in Twitter and disclosed the individual-level diurnal and seasonal
mood rhythms in cultures across the globe. The population’s mood status
was also used to conduct the political forecasting [19]. About the emotion
correlation, recent studies [4, 6] show that happiness is assortative in Twit-
ter. While other negative emotions’ correlations ar e not considered in these
studies and how the local structure affect the sentiment influence is also not
fully investigated. We try to focus more on the difference of correlation be-
tween different sentiments and probe deeper into the relation between local
structure and emotion correlation. We conjecture that emotion play a sig-
nificant role in information contagion, especially the negative emotions [5].
Because of this, understanding the corr elat ion difference could shed light on
the origin of abnormal event propagation in online social media and provide
many inspirations for modeling sentiment influence.
4

3. Methods
In this section, the methodology of t he present paper would be depicted.
First, we introduce the collection of the tweets from Weibo and the construc-
tion of the interaction network. Then the classifier we employ to mine the
sentiment from tweets is reported. Thirdly we define two kinds of emotion
correlations for connected users in the interaction network.
3.1. Weibo Dataset
As pointed out in [4], the following relationship in Twitter-like social
networks does not stand for the social interaction, while if two users reply,
retweet or mention each other in their tweets for certain times, the online so-
cial tie between t hem is sufficient to present an alternative means of deriving
a conventional social network [6]. So here we construct an interaction net-
work from the tweets we crawled from Weibo during April 2010 to September
2010, where interaction means the number that two users r etweet or mention
each other is larger than a threshold T . From around 70 million tweets a nd
200,000 users we crawled, an undirected but weighted graph G(V, E, T ) is
constructed, in which V is the set of users, E represents the set of interactive
links among V and T is the minimum number of interactions on each link.
For each link in E, its weight is the sum o f retweet or mention times between
its two ends in the specified time period. Specifically, to exclude occasional
users that are not truly involved in the Weibo social network, we only reserve
those active users in our interaction network that posted more than one tweet
every two days on average over the six months. And to guarantee the valid-
ity of users’ social interaction, if the number o f two users retweet or mention
each other is less t ha n T , we would omit the connection between them. As
shown in Figure 1, by tuning T we can obt ain networks of different scales.
Generally we set T = 30 and then the interaction network G contains 9868
nodes a nd 19517 links. We also make our entire data set publicly availa ble
4
.
3.2. Emotion classification
In this paper, t he emotion is divided into four classes, including anger,
sadness, joy and disgust. We employ the bayesian classifier developed in per-
vious work [15]. In this method, we use the emoticon, which is pervasively
used in Weibo, to label the sentiment of the tweets. At the first stage, 95
4
http://ipv6.nlsde.buaa.edu.cn/zhaojichang/emotionspread.tar.gz
5

Figure 1: The number of nodes or edges varies for different interaction thresh-
old T. In the following pa rt o f the present work, we set T = 30 to extract a
large enough network with convincing interaction strength.
(a) An interaction network. (b) Node colored by emotions.
Figure 2: The giant connected cluster of a network sample with T = 30.
(a) is the network structure, in which each node stands for a user and the
link between two users represents the interaction between them. Based on
this topology, we color each node by its emotion, i.e., the sentiment with the
maximum tweets published by t his node in the sampling period. In (b), the
red stands for anger, the green represents joy, the blue stands for sadness
and the black represents disgust. The regions of same color indicate that
closely connected nodes share the same sentiment.
6

frequently used emoticons are manually labelled by different sentiments and
then if a tweet only contains the emoticons of a certain sentiment, it would
be labeled with this sentiment. From around 70 million tweets, 3.5 million
tweets with valide emoticons are extracted and labeled. Using this data set
as a training corpus, a simple but fast bayesian classifier is built in the sec-
ond stage to mine the sentiment of the tweets without emoticons, which are
about 95% in Weibo. The averaged precision of this classifier is convincing
and particularly the large amount of tweets we employ in the experiment
can guarantee its accuracy further. Based on this framework, we demon-
strate a sampled snapshot of interaction network with T = 30. As shown
in Figure 2b, in which each user is colored by it s emotion. We can roug hly
find that closely connected nodes generally share the same color, indicating
emotion correlations in Weibo network. Besides, different colors show dif-
ferent clusterings. For example, the color of red, which represents an g er,
shows more evident clustering. These preliminary findings inspire us that
different emotions might have different correlations and a deep investigation
is necessary.
3.3. Emotion correlation
Emotion correlation is a metric to quantify the strength of sentiment
influence between connected users. Fo r a fixed T , we first extract an inter-
action network G and all the tweets posted by the nodes in G. Then by
employing the classifier established in the former section, the tweets for each
user is divided into four categories, in which f
1
, f
2
, f
3
and f
4
represent the
fraction of angry, joyful, sad and disgusting tweets, respectively. Hence we
can use emotion vector e
i
(f
i
1
, f
i
2
, f
i
3
, f
i
4
) to denote user i’s sentiment stat us.
Based on this, we define pairwise sentiment correlat ion as follows. Given a
certain hop distance h, we collect all user pairs with distance h from G. For
one of t he four emotions m(m = 1, 2, 3, 4) and a user pair (j, q), we put the
source user j’s f
j
m
into a sequence S
m
, and the target user q’s f
q
m
to a nother
sequence T
m
. Then the pairwise correlation could be calculated by Pearson
correlation as
C
m
p
=
1
l − 1
l
X
i=1
(
S
i
− hS
m
i
σ
S
m
)(
T
i
− hT
m
i
σ
T
m
),
where hS
m
i =
1
l
P
l
i=1
S
i
is the mean, σ
S
m
=
q
1
l−1
P
l
i=1
(S
i
− hS
m
i) is the
standard deviation and l is length of S
m
or T
m
. Or it can also be obtained
7

(a) Person correlation (b) Spearman correlation
Figure 3: Correlation for different emotions as the hot distance varies. Large
h means a pair o f users are far away from each other in the social network
we build. Here T = 30 is fixed.
from Spear ma n correlat ion as
C
m
s
= 1 −
6
P
l
i=1
d
2
i
l( l
2
− 1)
,
where d
i
is the rank difference between S
i
in S
m
and T
i
in T
m
. Intuitively
larger C
m
p
and C
m
s
both suggest a more positive correlation for sentiment m.
Based o n the dataset and classifier, interaction networks could be built
and tweets of each user in the network would be emotionally la belled. Using
the definition o f correlations, we can then present the comparison of emotion
correlations and the impact of local structures in the following section.
4. Results
First we compare the correlation of different emotions based on the gr aph
of T = 30, which ensures enough number of ties and users, and at the same
time g uarantees relatively high social tie strength. As shown in Figure 3,
both Pearson correlation and Spearman correlation indicate that different
sentiments have different correlations and anger has a surprisingly higher
correlation than other emotions. This suggests that anger could spread
quickly and broadly across the network because of its strong influence to
the neighborhoods in the scope of about three hops. Although the previous
studies [4, 6] show that happiness is assortative in online social networks,
8

(a) Person correlation (b) Spearman correlation
Figure 4: The emotion sequence is randomly shuffled to test the correlation
significance.
but Figure 3 further demonstrates that the correlation of anger is much
stronger than that of happiness. It means the information carrying angry
message might propagate very fast in the network and this phenomenon is
contrary to our intuition. While for sadness and disgust, they both share an
unexpected low cor relation even for small h. For instance, the correlation of
sadness is less than 0.15 as h = 1, which means sad status almost does not
affect the directly connected friends at all. The results are also consistent
with the previous findings that strength of the emotion correlation decreases
as h grows, especially after h > 6 [6]. In fact, as h > 3, the emotion corre-
lation becomes weak for all the sentiments, which means that the influence
of the sentiment in the social network is limited significantly by the social
distance. For example, for strong assortative emotions like anger and joy,
their correlations just fluctuate around 0 as h > 5.
In order to test the above correlation further, we also shuffle S
m
and
T
m
randomly for sent iment m and recalculate its correlation. As shown in
Figure 4, fo r the shuffled emotion sequence, there is no correlation existing
for all the sentiments. It indicates that the former correlation we get is
truly significant and for random pair of users in social network, there is no
emotion homophily. It further justifies that through social ties, the sentiment
indeed spreads between closely connected friends and different users could
influence their neighborhoods’ mood statuses because of the social bonds
between them.
Investigating to what extent the local structure, like tie strength and
9

(a) anger (b) joy
(c) sadness (d) disgust
Figure 5: Pearson correlations of different h fo r different networks extracted
by varying T. The case of h > 3 is not considered here because of the weak
sentiment correlation found in Figur 3.
Figure 6: Here T is fixed to 30. Because the network is relatively small, the
largest degree we get is o nly 30. The results just indicate when the degree is
small, all sentiments’ correlatio ns increase with node degrees.
10

node degree, could affect the emotion correlation is of importance for mod-
eling sentiment influence and propagation. As shown in Figure 5, we first
disclose how the interaction threshold T affects the sentiment influence. As
discussed in Section 3.1, larger T produces smaller networks but with closer
social relations and stro ng online interactions. It is also intuitive that fre-
quent interactions in online social networks are positively related with stro ng
social ties and convincing social bonds. Because of this, we can see in Figure 5
that for all the four emotions, their correlations inside two hops continue a
steady increasing trend with T ’s growth. Particularly for anger, its Pear-
son correlation could rise to around 0.52. For weakly correlated emotions
like sadness and disgust, alt hough t he correlat ion shows a slow growth for
h = 1 and h = 2, while the maximum value of the correlat ion is still lower
than 0.25. As h = 3, the increment of the sentiment influence is trivial,
esp ecially for sadness and disgust. It illustrates that the primary factor of
controlling emotion cor r elat ion is still the social distance and the social tie
strength just functions for close neighbors in the scope of two hops. Secondly,
we check the effect of users’ degrees to the sentiment influence. We select
a node i with degree k and then average its neighbors’ emotion vectors to
e
nei
i
(
1
k
P
j
f
j
1
,
1
k
P
j
f
j
2
,
1
k
P
j
f
j
3
,
1
k
P
j
f
j
4
), where j is an a rbitrary neighbor
of i. Through adding f
i
m
into S
m
and
1
k
P
j
f
j
m
into T
m
, we could get the
correlation of sentiment m for the users with degree k. As can be seen in
Figure 6, the sentiment correlation grows with k, especially for ang er and
joy, which illustrates tha t nodes with higher degrees in online social network
posses mor e significant emotional influence to their neighborhoods. This
finding is consistent with the conventional viewpoint that high degree nodes
in the social network own more social influence and social capital. Specifi-
cally, the cor r elat ion of anger and joy are almost same for very small degrees,
but later anger shows a significant jump for large degrees and enlarges the
gap as compared t o joy. As k rieses to 30, the corrleation of anger grows to
0.85. While the correlation of sadness and dis gust do not demonstrate an
obvious increasing t r end and just fluctuate around 0.2 or even lower. It is
worthy emphasizing that because the network size is small and we only have
maximum degree around 30, which is far below the Dunbar’s number[20].
We suspect that the correlation might stop rising if the degree is larger than
Dunbar’s number. The results of Spearman correlation are similar and not
reported here.
To sum up, different emotions have different correlations in the social me-
11

dia. Compared to other sentiments, anger has the most positive correlation,
which indicates its fast and broad propagation. Local structure can affect
the sentiment influence in near neighborhoods, from which we can learn that
tie strength and node degree both could enhance the sentiment influence,
esp ecially for anger and jo y, and their contr ibutions to s adness and disgust
are greatly limited. While high correlation of angry mood but weak influence
of sad status indeed require much more detailed explorations to disclose the
underlying reasons.
5. Empirical Explanation
With the continuous growth, online social medias in China like Weibo
have been becoming the primary channel of information exchange. In Weibo,
the messages do not only deliver the factual information but also propagate
the users’ opinions about the social event or individual affairs. Hence we
try to unravel the underlying reason o f why anger has a surprisingly high
correlation but the spread of sadness is weak from the view of keywords
the corresponding tweets present. For a certain emotio n m, we collect all
the retweeted tweets(usually contain phrase like “@” or “retweet”) with this
sentiment in a specified time period to combine into a long text document .
Focusing only on retweeted tweets could help reduce the impact of external
media and just consider social influence from the social ties in Weibo. Several
typical techniques are employed to mine the keywords or topic phrases from
the document s, which are reported in Figure 7. Based on the keywords or
topics we find, the real-world events or social issues could be summarized to
understand the sentiment influence in detail.
With respect to anger, we find two kinds of social events are apt to trig-
ger the angry mo od of users in Weibo. First one is the domestic social
problems like food security, government bribery and demolition for resettle-
ment. The “shrimp washing powder” which results in muscle degeneration
and the self-burning event in Fenggang Yihuang County of Jiangxi province
represent this category. These events reflect that people living in China a r e
dissatisfied about some aspects of the current society and this type of event
can spread quickly as the users want to show their sympathy to the victims
by retweeting tweets and criticizing the criminals or the government. Fr e-
quently appearing phrases like “government”, “bribery”, “demolition” and
so on are strongly related with these events. The second type is about the
diplomatic issues, such as the conflict between China and foreign countries.
12

(a) anger phrases (b) sadness phrases
Figure 7: The example Chinese keywords extracted for anger and sadness,
respectively. The to p 20 keywords are also translated int o English, which
could be fo und through http://goo.gl/7JIEgR.
For instances, in August 2010, United States and South Korea held a drill
on the Yellow Sea, which locates in the east of China. In September 201 0,
the ship collision of China and Japa n also made users in Weibo extremely
rageful. Actually, these events could arouse patriotism and stimulate the
angry mood. Keywords like “Diaoyu Island”, “ship collision” and “Philip-
pines” show the popularity of these events at that time. To sum up, Weibo
is a convenient and ubiquitously channel for Chinese to share their concern
about the continuous social problems and diplomatic issues. Pushed by the
real-world events, these users tend to retweet tweets, express their anger and
hope to get resonance from neighborhoods in online so cial network. While
regarding to sadness, we find its strength of correlation is strongly a ff ected
by the real-world natural disasters like earthquake, as shown in Figure 7b.
Because the natural disaster happens occasionally and then the averaged cor-
relation of the sadness is very low and the strength of its correlation might
be highly fluctuated.
In summary, real-world society issues are easy to get attention f rom the
public and people tend to express their anger towards theses issues through
posting and retweeting tweets in online social media. The angry mood de-
livered thro ugh social t ies could boost the spread of the corresponding news
and speed up the formation of public opinion and collective behavior. This
can explain why the events related to social problems propagate extremely
fast in Weibo.
13

6. Discussion and Conclusion
Users with similar demographics have high probabilities to get connected
in both online and offline social networks. Recent studies reveal that even
the psychological states like happiness are assortative, which means the hap-
piness or well-being is strongly correlated between connected users in online
social medias like Twitter. Considering the oversimplification of the sen-
timent classification in the previous literature, we divide the emotion into
four catego ries and discuss their different correlations in details based on
the tweets collected from Weibo of China, and the data set has been public
ava ilable to research community.
Our results show that ang e r is more influential than other emotions like
joy, which indicates that the angry tweets can spread quickly and broadly
in the networ k. While out of our expectation, the correlation of sadn ess is
low. Through keywords and topics mining in retweeted angry tweets, we
find the public opinion towards social problems and diplomatic issues are
always angry and t his extreme mental status a lso boost the propagation of
the information in Weibo. This might be the origin of large scale online
collective behavior in Weib o about society problems such as food security
and demolition for resettlement in recent year s. We conjecture that anger
plays a non-ignor able role in massive propagations of the negative news about
the society, which are always hot trends in today’s internet of China.
Besides, we also investigate the affect of local structure to the emotion
correlation in online social media, which is not fully probed before. We find
that for a pair of users the emotion correlation is stronger if mo r e interac-
tions happen between them. We also disclose that the a node’s degree could
significant ly enhance the sentiment influence to its neighborho od, especially
for anger and joy. These findings could shed light on modeling sentiment
influence and spread in social networ ks.
7. Acknowledgements
This work was partially supported by the fund of the State Key Labora-
tory of Software Development Environment under Grant SKLSDE-2011ZX-
02, the Research Fund for the Doctoral Program of Higher Education of
China under Grant 20111102110019, and the National 863 Program under
Grant 2012AA01 1005. JZ and YC both thank t he Innova tion Foundation of
BUAA for PhD Graduates.
14

References
[1] M. Miller, S.-L. Lynn, C. James, M, Birds of a feather: Homophily in
social networks, Annual Review of Sociology 27 (2001) 415– 444.
[2] A. Mislove, M. Marcon, K. P. Gummadi, P. Druschel, B. Bhattacharjee,
Measurement and analysis o f online social networks, in: the 7th ACM
SIGCOMM conference on Internet measurement, IMC ’07, 2007, pp.
29–42. doi:http://doi.acm.org/10.1145/1298306.1298311.
[3] D. Liben-Nowell, J. Kleinberg , The link prediction problem for so-
cial networks, in: the twelfth international conference on Informa-
tion and knowledge management, CIKM ’03, 2003, pp. 556–559.
doi:http://doi.acm.org/10.1145/956863.956972.
[4] J. Bollen, B. Gon¸calves, G. Ruan, H. Mao, Happiness is assortat ive in
online social networks, Artif. L ife 17 (3) (2011) 237–251.
[5] A. Chmiel, P. Sobkowicz, J. Sienkiewicz, G. Paltoglou, K. Buckley,
M. Thelwa ll, J. A. Holyst, Negative emotions boost user activity at
bbc forum, Physica A: Statistical Mechanics and it s Applications 390
(2011) 2936–2944.
[6] C. A. Bliss, I. M. Kloumann, K. D. Harris, C. M. Danforth, P. S. Dodds,
Twitter reciprocal reply networks exhibit assortativity with respect to
happiness, Journal of Computational Science 3 (2012) 388–397.
[7] A. L. Traud, E. D. Kelsic, P. J. Mucha, M. A. Porter, Comparing com-
munity structure to characteristics in online collegiate social networks,
ArXiv e-prints, arXiv0809.0690.
[8] D. Derks, A. H. Fischer, A. E. Bos, The role of emotion in computer-
mediated communication: A review, Computers in Human Behavior 24
(2008) 766–785.
[9] P. S. Dodds, C. M. Danforth, Measuring the happiness of large-scale
written expression: So ngs, blogs, and presidents, Journal of Happiness
Studies 11 (4) (2010) 441–456.
[10] P. S. Dodds, K . D. Harris, I. M. Kloumann, C. A. Bliss, C. M. Dan-
forth, Temporal pat terns of happiness and information in a global social
network: Hedonometrics and twitter, PLoS O NE 6.
15

[11] B. Pang, L. Lee, S. Vaithyanathan, Thumbs up?: sentiment classification
using machine learning techniques, in: EMNLP, 2002, pp. 79–86.
[12] R. Parikh, M. Movassate, Sentiment analysis of user-generated twitter
updates using various classification techniques, Technical report .
[13] J. Read, Using emoticons to reduce dependency in machine learning
techniques for sentiment classification, in: ACLstudent, 2005, pp. 43–
48.
[14] A. Go, R. Bhayani, L. Huang, Twitter sentiment classification using
distant supervision, Technical report, Stanford Digital Library Tech-
nologies Project.
[15] J. Zhao, L. Dong , J. Wu, K. Xu, Moodlens: an emoticon-based sentiment
analysis system for chinese tweets, in: KDD ’12, 2012, pp. 1528–1531.
[16] J. Bollen, A. Pepe, H. Mao, Modeling public mood and emotion: Twitter
sentiment and socio-economic phenomena, in: Fifth ICWSM, 2011.
[17] J. Bollen, H. Mao, X. Zeng, Twitter mood predicts the stock mar ket,
Journal of Computational Science 2 (1) (2011) 1–8.
[18] S. A. Golder, M. W. Macy, Diurnal and seasonal mood vary with work,
sleep, and day length across diverse cultures, Science 333 (6051) (2011)
1878–1881.
[19] M. Marchetti-Bowick, N. Chamb ers, Learning for microblogs with dis-
tant supervision: Political forecasting with twitter, in: 13th EACL,
2012, pp. 603–612.
[20] R. Dunbar, Grooming, Gossip, and the Evolution of L anguage, Harvard
University Press, Cambridge, MA, 1998.
16