Mood Lens: An Emoticon-Based sentiment Analysis System for Chinese Tweets Jichang Zhao Junjie Wul Ke xut zhaojichanganlsde.buaaedu.cndonglixp@gmail.comwujjabuaa.edu.cnkexu@nlsde.buaaedu.cn State Key Lab of Software Development Environment, Beihang University tBeijing Key Laboratory of Emergency Support Simulation Technologies for City Operations, School of Economics and Management, Beihang University Corresponding author ABSTRACT General terms Recent years have witnessed the explosive growth of online Measurement, Experimentation social media.Weibo,a Twitter-like online social network in China has attracted more than 300 million users in less than three years, with more than 1000 tweets generated in every Keywords econd. These tweets not only convey the factual informa- Sentiment Analysis; Chinese Short Text; Online Social Me- tion. but also reflect the emotional states of the authors dia: Weibo which are very important for understanding user behaviors. However, a tweet in Weibo is extremely short and the words 1. INTRODUCTION it contains evolve extraordinarily fast. Moreover. the Chi- se corpus of sentiments is still very small, which prevents The development of online social networks has attracted the conventional keyword-based methods from being used enormous Internet users in this decade. They are becoming In light of this, we build a system called MoodLens, which he mainstream online social media for information shar- to our best know ledge is the first system for sentiment anal ing.Twitter(www.twitter.com),amicroblogsitelaunched in 2006, has over 300 million registered users, with over 140 sis of Chinese tweets in Weibo. In MoodLens, 95 emoticons million microblog posts, known as tweets, being published are mapped into four categories of sentiments, i.e. angry disgusting, joyful, and sad, which serve as the class label everydayInChina,Weibo(www.weibo.com),aTwitter of tweets. We then collect over 3.5 million labeled tweet like service launched in 2009. has accumulated more than s the corpus and train a fast Naive Bayes classifier, with 300 millions users in less than three years. Every second an empirical precision of 64. 3%. MoodLens also more than 1000 Chinese tweets are posted in Weibo n incremental learning method to tackle the problem of Each user in the network can be viewed as a social set which publishes and propagates the information through the the sentiment shift and the generation of new words. Using tweets. Therefore, the huge amount of tweets convey com- ral interesting temporal and spatial patterns are observed. plicated signals of the authors and the real-world events, in Also, sentiment variations are well-captured by MoodLens which the sentiment is an essential part. In 3, the authors to effectively detect abnormal events in China. Finally, by argued that the events in the social, political and cultural using the highly efficient Naive Bayes classifier, MoodLens is fields did have a significant effect on the users'mood, which ofMoodlenscanbefoundathttp://goo.gl8ddesiscouldbedetectedbytheirtweetsItwasalsoclaimedin[2] capable of online real-time sentiment monitoring that the stock market even could be predicted by the senti ment analysis of the Twitter users Disclosing the emotions in tweets therefore plays a key Categories and Subject Descriptors role in understanding the user behaviors in social media However, both Twitter and Weibo only allow users to post H 2.8 Database Management: Database Application messages up to 140 characters, which makes the tweets ex- Data Mining: H 3.3 Information Search and Retrieval tremely short, and the sentiment analysis therefore becomes Text Mining;J. 4 Social and Behavioral Sciences: [Mis- a very challenging task. In particular, few works have been ellaneous done to reveal how to perform sentiment analysis for Chinese In light of this, we propose a system called MoodLens to perform the sentiment analysis for Chinese Weibo. The main Permission to make digital or hard copies of all or part of this work for contributions lie in the following aspects:(a)Moodlens er personal or classroom use is granted without fee provided that copies are ploys an emoticon-based method for sentiment classification, not made or distributed for profit or commercial advantage and that copies which helps to tackle the longstanding sparsity problem of bear this notice and the full citation on the first page. To copy otherwise, to hort texts;(b) MoodLens can detect four types of senti- nents: angry, disgusting, joyful, and sad, which goes be- KDD12 12-16.2012,Bej yond the traditional binary sentiment(positive vs. negative) Copyright2012ACM978-14503-1462-6/2/08….s500 analysis studies, and is crucial for unveiling the abundant
MoodLens: An Emoticon-Based Sentiment Analysis System for Chinese Tweets Jichang Zhao∗ Li Dong∗ Junjie Wu† Ke Xu∗‡ zhaojichang@nlsde.buaa.edu.cn donglixp@gmail.com wujj@buaa.edu.cn kexu@nlsde.buaa.edu.cn ∗State Key Lab of Software Development Environment, Beihang University †Beijing Key Laboratory of Emergency Support Simulation Technologies for City Operations, School of Economics and Management, Beihang University ‡Corresponding author ABSTRACT Recent years have witnessed the explosive growth of online social media. Weibo, a Twitter-like online social network in China, has attracted more than 300 million users in less than three years, with more than 1000 tweets generated in every second. These tweets not only convey the factual information, but also reflect the emotional states of the authors, which are very important for understanding user behaviors. However, a tweet in Weibo is extremely short and the words it contains evolve extraordinarily fast. Moreover, the Chinese corpus of sentiments is still very small, which prevents the conventional keyword-based methods from being used. In light of this, we build a system called MoodLens, which to our best knowledge is the first system for sentiment analysis of Chinese tweets in Weibo. In MoodLens, 95 emoticons are mapped into four categories of sentiments, i.e. angry, disgusting, joyful, and sad, which serve as the class labels of tweets. We then collect over 3.5 million labeled tweets as the corpus and train a fast Na¨ıve Bayes classifier, with an empirical precision of 64.3%. MoodLens also implements an incremental learning method to tackle the problem of the sentiment shift and the generation of new words. Using MoodLens for real-time tweets obtained from Weibo, several interesting temporal and spatial patterns are observed. Also, sentiment variations are well-captured by MoodLens to effectively detect abnormal events in China. Finally, by using the highly efficient Na¨ıve Bayes classifier, MoodLens is capable of online real-time sentiment monitoring. The demo of MoodLens can be found at http://goo.gl/8DQ65. Categories and Subject Descriptors H.2.8 [Database Management]: Database Applications— Data Mining; H.3.3 [Information Search and Retrieval]: [Text Mining]; J.4 [Social and Behavioral Sciences]: [Miscellaneous] Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. KDD’12, August 12–16, 2012, Beijing, China. Copyright 2012 ACM 978-1-4503-1462-6 /12/08 ...$5.00. General Terms Measurement, Experimentation Keywords Sentiment Analysis; Chinese Short Text; Online Social Media; Weibo 1. INTRODUCTION The development of online social networks has attracted enormous Internet users in this decade. They are becoming the mainstream online social media for information sharing. Twitter (www.twitter.com), a microblog site launched in 2006, has over 300 million registered users, with over 140 million microblog posts, known as tweets, being published every day. In China, Weibo (www.weibo.com), a Twitterlike service launched in 2009, has accumulated more than 300 millions users in less than three years. Every second, more than 1000 Chinese tweets are posted in Weibo. Each user in the network can be viewed as a social sensor, which publishes and propagates the information through the tweets. Therefore, the huge amount of tweets convey complicated signals of the authors and the real-world events, in which the sentiment is an essential part. In [3], the authors argued that the events in the social, political and cultural fields did have a significant effect on the users’ mood, which could be detected by their tweets. It was also claimed in [2] that the stock market even could be predicted by the sentiment analysis of the Twitter users. Disclosing the emotions in tweets therefore plays a key role in understanding the user behaviors in social media. However, both Twitter and Weibo only allow users to post messages up to 140 characters, which makes the tweets extremely short, and the sentiment analysis therefore becomes a very challenging task. In particular, few works have been done to reveal how to perform sentiment analysis for Chinese tweets in Weibo. In light of this, we propose a system called MoodLens to perform the sentiment analysis for Chinese Weibo. The main contributions lie in the following aspects: (a) Moodlens employs an emoticon-based method for sentiment classification, which helps to tackle the longstanding sparsity problem of short texts; (b) MoodLens can detect four types of sentiments: angry, disgusting, joyful, and sad, which goes beyond the traditional binary sentiment (positive vs. negative) analysis studies, and is crucial for unveiling the abundant
entiments contained in tweets; (c) MoodLens implements be obtained as c"(t)= arg max, P(c)Il P(wi llc),where an incremental learning scheme to deal with the problems (ci) is the prior probability of cj f the sentiment shift of words and the generation of new In order to validate the performance of the classifier, the rds;(d) MoodLens is capable of real-time tweet process- set of labeled tweets is divided into two sets randomly, in g and classification, and therefore can serve as a real-time cluding training set, denoted as Ttrain and testing set, de- abnormal event monitoring system. The demo of MoodLens noted as Ttest. The fraction of Ttrain, i.e., the fraction of t isnowavailableathttp://goo.gl/8dq65 training data, is denoted as ft= train. In Ttrain, the set of tweets labeled as ci is denoted as Ttrain, similarly, the tweets 2. EMOTICON-BASED METHOD in Ttest of c, is denoted as Ttest. In the testing set, the cor- We have noticed that the graphical emoticons are pop- rectly predicted tweets of c, is denoted as P3. From these ular in Weibo. In recent work [1], it has been found that definitions, we mainly employ three metrics in this paper to the graphical emoticons can convey strong sentiment. They describe the effectiveness of the classifier, which are listed their mood when post the tweet. as follows. Precision is defined as p= Hence. we could treat these emoticons as sentiment label f the tweets. In fact. it is a kind of crowdsourcing, i.e., the is defined as r= iEj=1 r.. F-measure is defined sers label the tweet with emoticons to express their emo- f=2pr/(p+r) tion themselves. Because of this, categorizing the emoticons In this demo, we use a standard bag of words as the fea- into different sentiments would make the tweets divided into ture, set fe=0.9, P(c)=0.25 and get a Naive Bayes clas- fferent emotion classes. Among over 1000 emoticons, we sifier, its precision is 64.3%, recall is 53.3% and F-measure manually select 95 ones as the sentiment labels(denoted as is58.3% E)and divide them into four different sentiment categories We also present a simple incremental learning approach including angry, disgusting, joyful and sadness. As show in to complement the original Naive Bayes classifier. Here Figure 1, there are 9 emoticons in angry, 14 emoticons in we could assume the tweets in Weibo is a stream, in which disgusting, 50 emoticons in joyful and 22 emoticons in sad there is a fraction of tweets(denoted u)are sentimentally respectively. labeled, then these labeled tweets could be used to update the prior probability of words. To verify the effectiveness of the method, the following experiment is performed. we Sentiment #Emoticons Typical emoticons randomly shuffle T and divide it into 50 pieces of same size. Then we use the first piece as the training set and obtain an Angry initial classifier. For the other 49 pieces, we treat them the tweet stream, which means they enter into the classifier Disgusting1。 one by one, and in each of them, there is a fraction(u) of tweets are randomly selected as labeled tweets and could be Joyful 50 used to update the classifier. As shown in Figure 2, as the index of pieces, denoted as s, grows, the p, r and f of the Sad 回囹 classifier indeed grows. Particularly, higher u means a larger fraction of labeled tweets are used to update the classifier, and then the more updates produce more improvement Figure 1: Sentiment categories and the typical emoticons in each class From Dec. 2010 to Feb. 2011. MoodLens has collected more than 70 million tweets from Weibo, We extract over 3.5 million tweets that contains emoticons in e as the la beled tweets set, denoted as T. It indicates that in Weibo (a)Precision Recall there is nearly 5% of the tweets labeled by the sentiment (c)F-measure emoticons. Finally we obtain 569, 229 angry tweets, 290, 444 weets. These tweets could be used to as a initial sentiment 0, 0.01, 0.05, 0.1 from bottom to top, respectively corpus for Weibo. For each tweet t in T, MoodLens converts it into a sequence of words wi, where w; is a word and In summary, MoodLens employs Naive Bayes classifier is its position in t. with incremental learning to predict the detailed sentiment of the tweets. For other solutions, like Liblinear [4, which nB)to build the classifier, which consumes little training consumes much more training time while gain less than 5%o time and predicts the category fast. From the labeled tweets improvement of precision. Moreover, it is also hard to in- we could obtain the word wi's prior probability of belonging corporate incremental learning approaches into it to the sentiment category c, is P(w ci)=="2u where j= 1, 2, 3 or 4, n(wi) is the times that w app 3. APPLICATIONS in all the tweets in the c, and Laplace smoothing Data Collection Weibo has published its APIs since is used to avoid the prol zero probability. Then we 2010 and through these APIs ould establish the naive lassifier as follows, for a un- tweets and some basic demographic attributes of the users labeled tweet t with word sequence wi, its category could We build a Weibo application named"Are you happy?! and
sentiments contained in tweets; (c) MoodLens implements an incremental learning scheme to deal with the problems of the sentiment shift of words and the generation of new words; (d) MoodLens is capable of real-time tweet processing and classification, and therefore can serve as a real-time abnormal event monitoring system. The demo of MoodLens is now available at http://goo.gl/8DQ65. 2. EMOTICON-BASED METHOD We have noticed that the graphical emoticons are popular in Weibo. In recent work [1], it has been found that the graphical emoticons can convey strong sentiment. They help the users to express their mood when post the tweet. Hence, we could treat these emoticons as sentiment labels of the tweets. In fact, it is a kind of crowdsourcing, i.e., the users label the tweet with emoticons to express their emotion themselves. Because of this, categorizing the emoticons into different sentiments would make the tweets divided into different emotion classes. Among over 1000 emoticons, we manually select 95 ones as the sentiment labels(denoted as E) and divide them into four different sentiment categories, including angry, disgusting, joyful and sadness. As show in Figure 1, there are 9 emoticons in angry, 14 emoticons in disgusting, 50 emoticons in joyful and 22 emoticons in sad, respectively. Figure 1: Sentiment categories and the typical emoticons in each class. From Dec. 2010 to Feb. 2011, MoodLens has collected more than 70 million tweets from Weibo. We extract over 3.5 million tweets that contains emoticons in E as the labeled tweets set, denoted as T. It indicates that in Weibo, there is nearly 5% of the tweets labeled by the sentiment emoticons. Finally we obtain 569,229 angry tweets, 290,444 disgusting tweets, 2,218,779 joyful tweets and 607,715 sad tweets. These tweets could be used to as a initial sentiment corpus for Weibo. For each tweet t in T, MoodLens converts it into a sequence of words {wi}, where wi is a word and i is its position in t. In MoodLens, we employ the simple method of Na¨ıve Bayes (NB) to build the classifier, which consumes little training time and predicts the category fast. From the labeled tweets, we could obtain the word wi’s prior probability of belonging to the sentiment category cj is P(wi k cj ) = n cj (wi)+1 P q (n cj (wq)+1) , where j = 1, 2, 3 or 4, n cj (wi) is the times that wi appears in all the tweets in the category cj and Laplace smoothing is used to avoid the problem of zero probability. Then we could establish the Na¨ıve Bayes classifier as follows, for a unlabeled tweet t with word sequence {wi}, its category could be obtained as c ∗ (t) = arg maxj P(cj )ΠiP(wi k cj ), where P(cj ) is the prior probability of cj . In order to validate the performance of the classifier, the set of labeled tweets is divided into two sets randomly, including training set, denoted as Ttrain and testing set, denoted as Ttest. The fraction of Ttrain, i.e., the fraction of the training data, is denoted as ft = |Ttrain| |T | . In Ttrain, the set of tweets labeled as cj is denoted as T cj train, similarly, the tweets in Ttest of cj is denoted as T cj test. In the testing set, the correctly predicted tweets of cj is denoted as P cj . From these definitions, we mainly employ three metrics in this paper to describe the effectiveness of the classifier, which are listed as follows. Precision is defined as p = P4 j=1 |P cj | |Ttest| . Recall is defined as r = 1 4 P4 j=1 |P cj | |T cj test| . F-measure is defined as f = 2pr/(p + r). In this demo, we use a standard bag of words as the feature, set ft = 0.9, P(cj ) = 0.25 and get a Na¨ıve Bayes classifier, its precision is 64.3%, recall is 53.3% and F-measure is 58.3%. We also present a simple incremental learning approach to complement the original Na¨ıve Bayes classifier. Here, we could assume the tweets in Weibo is a stream, in which there is a fraction of tweets (denoted u) are sentimentally labeled, then these labeled tweets could be used to update the prior probability of words. To verify the effectiveness of the method, the following experiment is performed. we randomly shuffle T and divide it into 50 pieces of same size. Then we use the first piece as the training set and obtain an initial classifier. For the other 49 pieces, we treat them as the tweet stream, which means they enter into the classifier one by one, and in each of them, there is a fraction(u) of tweets are randomly selected as labeled tweets and could be used to update the classifier. As shown in Figure 2, as the index of pieces, denoted as s, grows, the p, r and f of the classifier indeed grows. Particularly, higher u means a larger fraction of labeled tweets are used to update the classifier, and then the more updates produce more improvements. 0 5 10 15 20 25 30 35 40 45 50 0.65 0.66 0.67 0.68 0.69 0.7 0.71 s p (a) Precision 0 5 10 15 20 25 30 35 40 45 50 0.46 0.47 0.48 0.49 0.5 0.51 0.52 0.53 0.54 s r (b) Recall 0 5 10 15 20 25 30 35 40 45 50 0.54 0.55 0.56 0.57 0.58 0.59 0.6 0.61 s f (c) F-measure Figure 2: Experiments of incremental learning, u = 0, 0.01, 0.05, 0.1 from bottom to top, respectively. In summary, MoodLens employs Na¨ıve Bayes classifier with incremental learning to predict the detailed sentiment of the tweets. For other solutions, like Liblinear [4], which consumes much more training time while gain less than 5% improvement of precision. Moreover, it is also hard to incorporate incremental learning approaches into it. 3. APPLICATIONS Data Collection Weibo has published its APIs since 2010 and through these APIs, it is easy to obtain the public tweets and some basic demographic attributes of the users. We build a Weibo application named “Are you happy?!” and
Weibo grant us the opportunity to access its APIs with the is the most joyful day, Mar. ll is the saddest day, Jul. 13 application-level. For the limitation of requesting the aPI, is the most disgusting day and Jul. 24 is the angriest day it is necessary to select some probes from Weibo and then (refer to 6 for the link). MoodLens also draws the distri. collect data from them. From a large-scale user-pool we bution graph of everyday sentiment for each region of China llected before 2011. which contains more than 2.2 million and show how the sentiment evolves day by day dynamically users,MoodLens randomly selects 6, 800 active users, 200 (refer to 7 for the link users for each province or region in China. Here the "active Abnormal Event Detection Intuitively, abnormal events users"means these users should be true users but not spam. in the real world would definitely affect the people's emotion A simple filtering rule is used to filter them out, which and then the mental change would be reflected by the tweets hat MoodLens only chooses the user with more than 200 people publish. The basic idea of the detection method is but less than 3, 000 followers and has published less than first to find the turning point in the variation of the sen- 3.000 tweet timent and then to extract information of event from the tweets. MoodLens defines a sequence of fraction for the sen- timent ci as S '), where t is the observing time, its unit is likely to be a day or an hour. Assuming we observe the variation of the sentiment from t= ti to t= t2, then the averaged fraction tweets in ci could be defined as ◆◆鲁◆4◆ (St,_t)= (a)Hourly pattern b)Weekly pattern here At= t2-ti is the time window of observation. Hence MoodLens could get the sequence of relative variation for c V3= Then MoodLens defines the sequence of sentiment variation ing order and the top-k t is selected as the outlier time points, denoted as (t1, t2,, t). For each ti, MoodLens could extract the tweets posted at that time and perform the in- formation extraction for the event. Here MoodLens employs Figure 3: Temporal sentiment pattern the simplest way, the top 5 bi-gram terms of high frequency would be extracted to depict the event happened. We per entiment patterns The hou ttern of the senti- form this method on the data set of 2011 and find it coule ment is showed in Figure 3(a). It could be found that the detect almost all the abnormal events happened during the time from 6: 00 AM to 8: 00 AM is the saddest moment, which whole year. As shown in Figure 4, we mark the top-10 is different from the recent study from Twitter 5. In their events detected from A to J. In these detected events, A, D data set, they found people are likely to be positive at early and E correspond to the event of bullet train crash. C and morning. While for Weibo, this period is also the angriest B correspond to the fact that Japan was hit by a magnitude moment for most of the users during a day. Surprised by this 9.0 earthquake, while corresponds to the news that people difference, we carefully investigate the tweets published inin China rushed to purchase the salt because of rumors. It Weibo from 6: 00AM to 8: 00AM and extract the commonly should be noted that for j, the fraction of angry is larger sed words. And the results show that the "sad"mood is than C and B. F corresponds to New Years' Eve of 2011 generally caused by the followed reasons. First, some people G and I correspond to Spring Festival of 2011. H corre- hate to get up early, but they have to. Second, some users sponds to the death of Steven Jobs. It is also interesting do not want to work at this time. Third, some ones might that different from C, B,J, A, D and E, for H, although have nightmare in the last night and have bad sleep. After the fraction of sad is high but the fraction of angry is low 10: 00 AM, users of Weibo seems to become more and more It indicates that detailed negative sentiments are useful for ful gradually and the fraction of joyful tweets reaches the analyzing the essence of the abnormal event eak at 20: 00 PM. The weekly pattern of th e sentiment Real-time Sentiment Monitoring The NB classifier showed in Figure 3(b). As can be seen, people seems to with incremental learning is speedy enough for the real-time ecome happier since Friday, the joyful reaches peak at sa sentiment analysis of the tweets in Weibo. Through the API urday and then the mood of joy goes down as sunday begi ovided by Weibo, MoodLens could obtain the most recent As shown in Figure 3(c)is the monthly pattern of the senti- 400 public tweets every minute and these tweets could be ment. There are several outliers. For instance, in March of analyzed in less than one second. In order to guarantee the 2011. it shows the f Weibo are sad and angry, it might statistical significance, we set the cycle of collecting tweets a be caused by the earthquake in Japan and the rumor that 30 minutes, which means MoodLens would download nearly the iodized salt in China is also nuclear polluted. Another 12, 000 tweets in one monitoring cycle. Then these twee one is in July, the fraction of angry reaches the peak, it is would be categorized into different sentiment classes in less mainly related with the accident of the bullet train. We also than 1 minute. As shown in Figure 5, we present a sample find the days of extreme sentiment in 2011. Namely, Jan. 1 cycle from the real-time monitoring, which starts from 19: 30
Weibo grant us the opportunity to access its APIs with the application-level. For the limitation of requesting the API, it is necessary to select some probes from Weibo and then collect data from them. From a large-scale user-pool we collected before 2011, which contains more than 2.2 million users, MoodLens randomly selects 6,800 active users, 200 users for each province or region in China. Here the “active users” means these users should be true users but not spam. A simple filtering rule is used to filter them out, which is that MoodLens only chooses the user with more than 200 but less than 3,000 followers and has published less than 3,000 tweets. 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 0.1 0.15 0.2 0.25 0.3 0.35 0.4 Hour fraction Angry Disgusting Joyful Sad (a) Hourly pattern. Mon Tue Wed Thu Fir Sat Sun 0.1 0.15 0.2 0.25 0.3 0.35 0.4 fraction Angry Disgusting Joyful Sad (b) Weekly pattern. 0 2 4 6 8 10 12 0.1 0.15 0.2 0.25 0.3 0.35 0.4 Month fraction Angry Disgusting Joyful Sad (c) Monthly pattern. Figure 3: Temporal sentiment patterns. Sentiment Patterns The hourly pattern of the sentiment is showed in Figure 3(a). It could be found that the time from 6:00 AM to 8:00 AM is the saddest moment, which is different from the recent study from Twitter [5]. In their data set, they found people are likely to be positive at early morning. While for Weibo, this period is also the angriest moment for most of the users during a day. Surprised by this difference, we carefully investigate the tweets published in Weibo from 6:00AM to 8:00AM and extract the commonly used words. And the results show that the “sad” mood is generally caused by the followed reasons. First, some people hate to get up early, but they have to. Second, some users do not want to work at this time. Third, some ones might have nightmare in the last night and have bad sleep. After 10:00 AM, users of Weibo seems to become more and more joyful gradually and the fraction of joyful tweets reaches the peak at 20:00 PM. The weekly pattern of the sentiment is showed in Figure 3(b). As can be seen, people seems to become happier since Friday, the joyful reaches peak at Saturday and then the mood of joy goes down as Sunday begins. As shown in Figure 3(c) is the monthly pattern of the sentiment. There are several outliers. For instance, in March of 2011, it shows the users of Weibo are sad and angry, it might be caused by the earthquake in Japan and the rumor that the iodized salt in China is also nuclear polluted. Another one is in July, the fraction of angry reaches the peak, it is mainly related with the accident of the bullet train. We also find the days of extreme sentiment in 2011. Namely, Jan. 1 is the most joyful day, Mar. 11 is the saddest day, Jul. 13 is the most disgusting day and Jul. 24 is the angriest day (refer to [6] for the link). MoodLens also draws the distribution graph of everyday sentiment for each region of China and show how the sentiment evolves day by day dynamically (refer to [7] for the link). Abnormal Event Detection Intuitively, abnormal events in the real world would definitely affect the people’s emotion, and then the mental change would be reflected by the tweets people publish. The basic idea of the detection method is first to find the turning point in the variation of the sentiment and then to extract information of event from the tweets. MoodLens defines a sequence of fraction for the sentiment cj as {S cj t }, where t is the observing time, its unit is likely to be a day or an hour. Assuming we observe the variation of the sentiment from t = t1 to t = t2, then the averaged fraction tweets in cj could be defined as hS cj t1→t2 i = 1 t2 − t1 Xt2 t=t1 S cj t , (1) where ∆t = t2−t1 is the time window of observation. Hence, MoodLens could get the sequence of relative variation for cj as V cj t = S cj t − hS cj t1→t2 i hS cj t1→t2 i . (2) Then MoodLens defines the sequence of sentiment variation as { P4 j=1 |V cj t |}. This sequence could be sorted in descending order and the top − k t is selected as the outlier time points, denoted as {t1, t2, .., tk}. For each ti, MoodLens could extract the tweets posted at that time and perform the information extraction for the event. Here MoodLens employs the simplest way, the top 5 bi-gram terms of high frequency would be extracted to depict the event happened. We perform this method on the data set of 2011 and find it could detect almost all the abnormal events happened during the whole year. As shown in Figure 4, we mark the top − 10 events detected from A to J. In these detected events, A,D and E correspond to the event of bullet train crash. C and B correspond to the fact that Japan was hit by a magnitude 9.0 earthquake, while J corresponds to the news that people in China rushed to purchase the salt because of rumors. It should be noted that for J, the fraction of angry is larger than C and B. F corresponds to New Years’ Eve of 2011. G and I correspond to Spring Festival of 2011. H corresponds to the death of Steven Jobs. It is also interesting that different from C, B, J, A, D and E, for H, although the fraction of sad is high but the fraction of angry is low. It indicates that detailed negative sentiments are useful for analyzing the essence of the abnormal event. Real-time Sentiment Monitoring The NB classifier with incremental learning is speedy enough for the real-time sentiment analysis of the tweets in Weibo. Through the API provided by Weibo, MoodLens could obtain the most recent 400 public tweets every minute and these tweets could be analyzed in less than one second. In order to guarantee the statistical significance, we set the cycle of collecting tweets as 30 minutes, which means MoodLens would download nearly 12,000 tweets in one monitoring cycle. Then these tweets would be categorized into different sentiment classes in less than 1 minute. As shown in Figure 5, we present a sample cycle from the real-time monitoring, which starts from 19:30
ry -Disgusting -Joyful--s NYWWWMAME H B 4 D XW 010203040506070809010011120130140150160170180190200210220230240250260270280290300310320330340350360370 Day Figure 4: Abnormal event detection of 2011. We also provide an interactive version on the following link: http://gana.nlsde.buaaedu.cn/hourly-happy/timeline_2011.html. sad. Moodlens is now online for temporal and spa- FAngry-Disgusting-Joyful--Sad tial sentiment pattern ry, abnormal events detection and illustration and al-time monitoring of senti- ment fluctuations 5. ACKNOWLEDGEMENTS This works was supported by the fund of the State Ke Laboratory of Software Development Environment(SKLSDE 2011ZX-02)and the Research Fund for the Doctoral Pro- gram of Higher Education of China(Grant No. 20111102110019 The third author was supported in part by the National Natural Science Foundation of China(NSFC) under Grants 71171007,70901002,and90924020 6. REFERENCES Figure a sample from Imon- Sho aoki and osamu Uchida. A Method for g online appl automatically generating the emotional vectors of gana. nlsde. buaa. edu. cn/hourly-happy/line- 36h.htmL. emoticons using weblog articles. ACACOS'1l, pages 132-136, Stevens Point, Wisconsin, USA, 201 2 Johan Bollen, Huina Mao, and Xiaojun Zeng. Twitter PM of Dec. 24. 2011 and ends to 7: 00 AM of Dec. 26. 2011 Mood predicts the stock market. Journal of the delay between two points is 30 minutes and 72 point omputational Science, 2(1): 1-8, 2011 re drew in all. As can be seen, the real-time event could 3 Johan Bollen, Alberto Pepe, and Huina Mao Modeling be indeed monitored. For instance, regarding to the time public mood and emotion: Twitter sentiment and marked by A. which is at 0: 00 PM of Dec. 24. the sentiment ocio-economic phenomena. 5th ICWSM, 2011 of joyful reaches peak because everyone is celebrating the 4 Rong-En Fan, Kai-Wei Chang, Cho-Jui Hsieh coming of the Christmas Day. While with respect to B, at ang-Rui Wang, and Chih-Jen Lin. Liblinear: A the time of 1: 00 AM of Dec. 26. the fraction of sad tweets library for large linear classification. J. Mach. Learn rises suddenly, the bi-gram terms reveal that it is caused by Res.9:1871-1874,June2008. the earthquake happened in Chengdu at that time. 5 Scott A. Golder and Michael W. Macy. Diurnal and seasonal mood vary with work, sleep, and daylength 4. CONCLUSION across diverse cultures. Science, 333(6051): 1878-1881 MoodLens is an online sentiment analysis system r Chi- September 2011 ese tweets in Weibo. It employs the emoticons for the gen-[6 ation of sentiment labels for tweets, and builds an incre- http://gana.nlsde.buaa.edu.cn/hourly-happy/line-cycl mental learning Naive Bayes classifier for the categorization 7] of four types of sentiments: angry, disgusting, joyful, and http://gana.nlsde.buaaedu.cn/hourly-happy/map-pie.html
0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200 210 220 230 240 250 260 270 280 290 300 310 320 330 340 350 360 370 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Day fraction Angry Disgusting Joyful Sad E F C A D H G I J B Figure 4: Abnormal event detection of 2011. We also provide an interactive version on the following link: http://gana.nlsde.buaa.edu.cn/hourly happy/timeline 2011.html. 0 10 20 30 40 50 60 70 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 time fraction Angry Disgusting Joyful Sad B A Figure 5: A sample from real-time monitoring. The online application could be accessed through the following link: gana.nlsde.buaa.edu.cn/hourly happy/line 36h.html. PM of Dec. 24, 2011 and ends to 7:00 AM of Dec. 26, 2011, the delay between two points is 30 minutes and 72 points are drew in all. As can be seen, the real-time event could be indeed monitored. For instance, regarding to the time marked by A, which is at 0:00 PM of Dec. 24, the sentiment of joyful reaches peak because everyone is celebrating the coming of the Christmas Day. While with respect to B, at the time of 1:00 AM of Dec. 26, the fraction of sad tweets rises suddenly, the bi-gram terms reveal that it is caused by the earthquake happened in Chengdu at that time. 4. CONCLUSION MoodLens is an online sentiment analysis system for Chinese tweets in Weibo. It employs the emoticons for the generation of sentiment labels for tweets, and builds an incremental learning Na¨ıve Bayes classifier for the categorization of four types of sentiments: angry, disgusting, joyful, and sad. MoodLens is now available online for temporal and spatial sentiment pattern discovery, abnormal events detection and illustration, and online real-time monitoring of sentiment fluctuations. 5. ACKNOWLEDGEMENTS This works was supported by the fund of the State Key Laboratory of Software Development Environment (SKLSDE- 2011ZX-02) and the Research Fund for the Doctoral Program of Higher Education of China (Grant No. 20111102110019). The third author was supported in part by the National Natural Science Foundation of China (NSFC) under Grants 71171007, 70901002, and 90924020. 6. REFERENCES [1] Sho Aoki and Osamu Uchida. A method for automatically generating the emotional vectors of emoticons using weblog articles. ACACOS’11, pages 132–136, Stevens Point, Wisconsin, USA, 2011. [2] Johan Bollen, Huina Mao, and Xiaojun Zeng. Twitter mood predicts the stock market. Journal of Computational Science, 2(1):1–8, 2011. [3] Johan Bollen, Alberto Pepe, and Huina Mao. Modeling public mood and emotion: Twitter sentiment and socio-economic phenomena. 5th ICWSM, 2011. [4] Rong-En Fan, Kai-Wei Chang, Cho-Jui Hsieh, Xiang-Rui Wang, and Chih-Jen Lin. Liblinear: A library for large linear classification. J. Mach. Learn. Res., 9:1871–1874, June 2008. [5] Scott A. Golder and Michael W. Macy. Diurnal and seasonal mood vary with work, sleep, and daylength across diverse cultures. Science, 333(6051):1878–1881, September 2011. [6] http://gana.nlsde.buaa.edu.cn/hourly happy/line cycle.html. [7] http://gana.nlsde.buaa.edu.cn/hourly happy/map pie.html
