helpfn

SPECIAL SECTION ON INTELLIGENT SENSING ON MOBILE AND SOCIAL MEDIA ANALYTICS

Received August 19, 2016, accepted September 3, 2016, date of publication September 15, 2016, date of current version March 15, 2017.

Digital Object Identifier 10.1109/ACCESS.2016.2607218

Sentiment Computing for the News Event Based on the Social Media Big Data DANDAN JIANG1, XIANGFENG LUO1,2, JUNYU XUAN3, AND ZHENG XU4 1School of Computer Engineering and Science, Shanghai University, Shanghai 200444, China 2Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100080, China 3Faculty of Engineering and IT, University of Technology Sydney, Ultimo, NSW 2007, Australia 4Third Research Institute of the Ministry of Public Security, Shanghai 200031, China

Corresponding author: X. Luo ([email protected])

This work was supported in part by the National Science Foundation of China under Grant 61471232 and Grant 91324005, in part by the Shanghai Science International Cooperation Project under Grant 16550720400, and in part by the Innovation Project of Institute of Computing Technology, Chinese Academy of Sciences.

ABSTRACT The explosive increasing of the social media data on the Web has created and promoted the development of the social media big data mining area welcomed by researchers from both academia and industry. The sentiment computing of news event is a significant component of the social media big data. It has also attracted a lot of researches, which could support many real-world applications, such as public opinion monitoring for governments and news recommendation for Websites. However, existing sentiment computing methods are mainly based on the standard emotion thesaurus or supervised methods, which are not scalable to the social media big data. Therefore, we propose an innovative method to do the sentiment computing for news events. More specially, based on the social media data (i.e., words and emoticons) of a news event, a word emotion association network (WEAN) is built to jointly express its semantic and emotion, which lays the foundation for the news event sentiment computation. Based on WEAN, a word emotion computation algorithm is proposed to obtain the initial words emotion, which are further refined through the standard emotion thesaurus. With the words emotion in hand, we can compute every sentence’s sentiment. Experimental results on real-world data sets demonstrate the excellent performance of the proposed method on the emotion computing for news events.

INDEX TERMS Text mining, sentiment computing, emotion classification, social media big data.

I. INTRODUCTION As the development of the web and the explosive increas- ing of people’s willingness to express and exchange their interests and opinions on the web, the social media data emerges, including but not limited to: facebook, twitter, weibo, and wechat. The social media data is considered as an instance of big data in the literature [1] due to the following reasons: (i) Volume The average number of microblogs to report a news event is at least 100,000; (ii) Variety Each microblog is composed by different mate- rials (such as, words and emoticons); (iii) Velocity The social media data is highly dynamic. For example, the new com- ing data in each day is more than 500TB in Facebook. (vi) Value The rich information hidden in the social media data is a perfect testing ground for the researchers in the big data area and also a powerful tool for the corpora- tions and governments to make specific decisions or global strategies.

News events, which is a significant component of the social media big data on the web, are news stories which have occurred in the society or on the web and are reported or discussed by a number of web pages on the web [2]. For example, MH 370 is a news event. After the occurrence of this web event, many people discuss it on the web through the social media. Browsing and discussing the news events has become the routine of daily life. Therefore, rich valuable information could be found through the news event analysis for the corporations and governments. Among lots of news event analysis, one of the most challenge tasks is the senti- ment computing for the news events, which aims to discover the emotion of the texts (i.e., microblogs) 1 from the users. However, the methods proposed almost classify texts into two

1Since the texts in social media are almost short texts and each short text can be seen as a sentence, we do not make the distinction on text/short text/sentence in this paper.

VOLUME 5, 2017 2169-3536 2016 IEEE. Translations and content mining are permitted for academic research only.

Personal use is also permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

2373

D. Jiang et al.: Sentiment Computing for the News Event Based on the Social Media Big Data

categories: positive and negative, which doesn’t conform the characteristic that the public sentiment is complex. Nowa- days, some researchers tend to compute the text multidimen- sional emotion. According to commonly-used emotion [3], it is a six-dimensional vector on: joy, love, surprise, fear, sad, anger (More details can be found in Section 3). For example, the emotion of a microblog about MH 370 from a user may be < 0.001,0.001,0.9,0.2,0.096,0.002 >, which means this user feels surprise and a little fear on this event.

The ability to discover the sentiment of a news event can be used in a variety of settings. For example, 1) public opinion monitoring. There are lots of information about a news event on the web, so accurately locating the information with apparent emotiveness (i.e., anger) could help the deci- sion makers understand the public concerns in a news event; 2) product feedback monitoring. After publishing a new prod- uct, a corporation usually wants to know the interests and concerns of its users about this product. Accurately locating the information with apparent emotiveness (i.e., surprise and anger) could help this corporation evaluate its former product design and make decision on the future design.

In these literatures, there are some state-of-the-art works have been developed for the affective computing for the document (More details can be found in Section 2). How- ever, existing works cannot be directly adopted for the news events due to the following drawbacks: 1) only focusing on using adjectives and adverbs for emotion classification, so the semantic of documents is overlooked; 2) there are two extremes: one highly relies on standard emotion the- saurus and the other completely abandons standard emotion thesaurus.

FIGURE 1. Framework of text sentiment computation.

Our innovative idea is to do the sentiment computing for a news event by the help of its semantic. As in Fig. 1, the main part of the news event sentiment computing task is the word emotion computation, which can be splitted into two procedures: word emotion computation through word emotion association network and word emotion refinement through standard sentiment thesaurus. For the first part of word emotion computation, a Word Emotion Association Network (WEAN) is built to jointly capture its semantic and emotion, which is the basis for both word and text emotion computations. Our assumption is that the words semantically associated will be more possible to have the similar emotions. An iterative process with its convergence proof is designed to optimize the emotional weights assignment for the links in WEAN. After this process, we can obtain the initial emotion

of words, but they may not be consistent with existing com- mon knowledge. For example, the word ‘happy’ should have a large value on joy, but it may obtain wrong emotion after the iterative process. So, in the second part of word emotion computation, we have designed a mechanism to refine the initial words emotion by incorporating the common prior knowledge: standard emotion thesaurus.

The following are the two major contributions of this paper:

1) A Word Emotion Association Network is built and utilized for word emotion computation, which captures both the semantic and emotion of a news event;

2) Standard emotion thesaurus is brought in for word emotion computation, which improves the accuracy of word emotion.

The remainder of this paper is organized as follows. Section 2 reviews related work; The proposed methods are presented in Section 3, which includes two parts: word emotion computation through word emotion association net- work and word emotion refinement through standard senti- ment thesaurus. Section 4 presents experimental setup and results on the real-world data in detail. Section 5 summarizes the findings and conclusions of this study with a discussion on future directions.

II. RELATED WORK Sentiment computation2 (also known as sentiment analysis) refers to use of natural language processing, or machine learn- ing methods to extract, identify, or characterize the sentiment context of a text. At the beginning of the study, researchers used documents as the data set and classified documents into two categories, one is the positive category and the other is negative category. As the social media grown, analyzing short texts’ sentiment becomes more and more popular. So the work on sentiment computation can be divided into two parts, sentiment computation (original) and short text sentiment computation.

A. SENTIMENT COMPUTATION At first, sentiment computation almost aims at dealing with long texts. Long text sentiment computation is a text clas- sification problem indeed, so the existing methods for text classification can all be used in this problem.

Pang et al. [4] used machine learning techniques to analyze the sentiment of film reviews, and they divided film reviews into two categories: positive sentiment and negative senti- ment. Three machine learning methods were taken in their experiments, and the result shows that using bag of words as features and using SVM as classifier has the best effect. After that, Cui et al. [5] made experiments on sentiment classification to prove that unigram was superior with not too many training corpus, and with the increasing of training

2There are also some works using ‘sentiment computing’ instead of ‘affective computing’ or ‘sentiment’ instead of ‘emotion’. They may have different implications in Psychology but normally with similar meaning in text mining area. So we do not make a distinction on two words in this paper.

2374 VOLUME 5, 2017

D. Jiang et al.: Sentiment Computing for the News Event Based on the Social Media Big Data

corpus, n-gram played a very important role. Kim et al. [6] computed sentence level sentiment with position feature and comment word feature. Taboada et al. [7] computed each document’s emotion score with the features of lexicon includ- ing emotion words and phrases about emotion orientation. Aue and Gamon [8] used semi-supervised method in learning sentiment with a small amount of sentences of even no labels. Blair-Goldensohn et al. [9] considered sentence fragments to recognized online users’ sentiment only using second person. Bao et al. [10] proposed a topic model named Emotion-Topic Model (ETM) to compute short texts’ sentiment of, whose dataset is Sina microblog. It added an intermediate layer into LDA means every sentiment is composed by several topics with different weight. Rao et al. [11] also presented topic models named Multi-label Supervised Topic Model (MSTM) and Sentiment Latent Topic Model (SLTM) to compute Sina microblog’s sentiment. Unlike Bao et al. [10], Rao et al. used words and emoticons in microblogs to train the models, and the results showed that MSTM and SLTM all performed better than ETM. Matthias Steinbauer et al. [12] constructed emotion lexicon according to film review, which is based on ten dimensions of emotion. Emotion lexicon includes emoti- cons and emotion scores. Emoticon role was divided into assuagement, conversion, addition and emphasis and then decided one tweet sentiment score based on emoticon role. Tang et al. [13] used deep learning to learn semantic repre- sentations of user and products for document level sentiment classification.

With the development of social media, the public prefer to present their opinion or sentiment by short texts rather than long documents. The methods dealing with long documents is not available for short text sentiment computation.

B. SHORT TEXT SENTIMENT COMPUTATION Recent years have witnessed the rapid development of social media platforms, such as Facebook, Twitter, Sina Microblog and so on. Read [14] generated positive Twitter corpus and negative Twitter corpus with appropriate emoticons. Go et al. [15] built a machine-learning classifier with fea- tures of Bigrams, unigrams and POS tags to divided Twitter corpus into two categories: positive and negative. And the most effective method was using NB with feature of bigram. Pak and Paroubek [16] used features of emoticons, POS and syntax to train classifier by generating Twitter public emotion corpus. Unlike Go, they classified tweets as posi- tive, negative and neutral. Zhang et al. [17] used supervised method to classify texts into positive and negative sentiment. Davidov et al. [18] presented a supervised approach with the features of hashtags, smileys, punctuations and fre- quent patterns to classify tweets into two categories, which was similar to a k-Nearest Neighbors algorithm (kNN). Bakliwal et al. [19] made eleven features train the SVM classifier, and used two different datasets to evaluate their approach. Jiang et al. [20] analyzed short text sentiment based on the target words of tweets and added the context information into the event text. Hu et al. [21] computed text

sentiment with the feature of emotion signals in social media. Saif et al. have done much to analyze Twitter sentiment. From reducing data sparseness of tweets [22] to removing stop words on Twitter sentiment effectiveness [23], and then presented SentiCircles, a lexicon-based approach to address Twitter sentiment analysis [24].

Most of the methods classify the texts into two categories: positive and negative, which is too simple to deal with sen- timent. Public sentiment is very complex that it should be classified into multidimensional sentiments.

III. PROPOSED METHOD The affective computing for the news events, which aims to obtain the emotion of its microblogs, consists of two parts: word emotion computation through Word Emotion Associ- ation Network (Section 3.1) and word emotion refinement through standard sentiment thesaurus (Section 3.2). Word emotions are the basis for the text emotion computation. In the following, we will successively introduce two parts in detail. The notations in this paper are listed in Table 1.

TABLE 1. Notations in this paper.

A. WORD EMOTION COMPUTATION THROUGH WORD EMOTION ASSOCIATION NETWORK People can publish views, opinions and attitudes for objects, individuals, events or topics on social media whenever and wherever. These views, opinions and attitudes contain users’ emotion. We call it social media text emotion. Up to now, there are many researches on social media text emotion, and many types of multidimensional emotion have been reported from different views. We use multidimensional emotion clas- sification published by; in which emotions are divided as love, joy, anger, sad, fear and surprise.

We propose a method to compute emotion of words in a news event. Given a news event, a word in different stages may have different emotion, so we need to compute the word emotion at a specific time. Due to the large scale of words number and the varied usage of words, comput- ing word emotion is extremely difficult. What is worse, with the development of the web, the appearance of new cyber words makes word emotion computation even harder.

VOLUME 5, 2017 2375

D. Jiang et al.: Sentiment Computing for the News Event Based on the Social Media Big Data

Fortunately, people use words express their emotion with emoticons in Microblog, and the position of emoticons can be anywhere: before a sentence, after a sentence or within a sentence. The number of default emoticons is limited, and their emotions are obvious. We can compute word emotion by the help of emoticons.

In this study, we design a questionnaire to gather opinions about emoticons’ six-dimensional sentiment. Twenty under- graduates who have the experience with using emoticons in their microblogs have joined this questionnaire. This step is prepared for computing the emotion of words.

FIGURE 2. An illustration of Word Emotion Association Network for ek .

Definition 1 (Word Emotion Association Network (WEAN)): Like association link network (ALN) [25], WEAN is constructed by words and their association rules, which can be represented by,

WEAN =< N,W > (1)

where N is a set of nodes and W is a set of weighted links belonging to N × N. In Fig. 2, each node denotes a word which can be noun, verb, adjective and out of vocabulary. When there is a link between two words, it means these two words have co-occurrence. Different from ALN, the weight of the link ωe

k

i,j shows the strength of the emotion relationship between two words, which is decided by the number of emoti- cons and the emoticons’ value. Emoticons come from all of the microblogs which appear with these two words together. ωe

k

i,j can be computed by,

ω ek i,j =

M∑ x=1

we k

emoticonx (2)

where i denotes word i, j denotes word j, ek is the k-th dimension of emotion, M is the emoticons number from all of the microblogs which contain both word i and j. we

k

emoticonx denotes the emotion value of emoticon x on ek, which are obtained through the above questionnaire. If word i and j do not have concurrence, then ωe

k

i,j = 0. Note that the

link emotion weight is symmetric, which means Fig. 2 is a bidirectional graph and ωe

k

i,j = ω ek j,i.

Before computing the emotion of words, it is necessary to do word segmentation first. Considering the novelty and mutability of web language, cyber words may have stronger emotion and express people’ emotion better than the ordi- nary words. The traditional word segmentation tools cannot segment popular cyber words effectively, which affects the computation of word emotion. So it is necessary to import new words when segmenting microblogs. In order to focus on the computing word emotion, we omit how to import new words and segment microblogs.

The part-of-speech of words includes nouns, verbs, adjec- tives, adverbs, quantifiers, preposition and so on. Among these, nouns, verbs, adjectives and cyber words should be the words that have stronger emotion than other words. Adverbs often strengthen or weaken emotion of verbs or adjectives to some extent. Therefore, word emotion computation is for nouns, verbs, adjectives and cyber words. We construct a word emotion association network (WEAN) in Definition 1 for the Microblogs whose ek is not null.

After WEAN is constructed, the next step is to compute word emotion of ek. The different scale and intension of word circumstance can affect word emotion. The larger of the scale and the stronger of the intension, the more intense of the word emotion. Based on these, we propose two assumptions: • Quantity Assumption: In WEAN, if one word has links with other words who have this emotion, the more links exist, the stronger emotion the word has. On the contrary, the less of the links, the weaker of the emotion;

FIGURE 3. Quantity Assumption example.

In Fig. 3, each node denotes a word that appears in emotion ek, such as love. N1...N6 are nodes we already know their emotion values, and suppose they have the same value. T1 and T2 are the words that we need to compute their emotion. T1 has two words linked with it, while T2 has four. According to quantity assumption, T2 has stronger emotion in love than T1.

• Intensity Assumption: Words can affect other nearby words through the links. The words who link with word kn can have different intensity of ek. The stronger of the word, the more emotion it transmits to other words. Hence, if one word links to many words whose ek are strong, the emotion ek of word will be strong too. In Fig. 4, N7...N10 and T3,T4 are nodes that appears in emotion ek. N7...N10 are nodes that already known emotion values, while N7 and N8 have stronger emotion in ek than N9 and N10, Where N7 and N8 are smaller than

2376 VOLUME 5, 2017

D. Jiang et al.: Sentiment Computing for the News Event Based on the Social Media Big Data

FIGURE 4. Intensity Assumption example.

N9 and N10. T3 and T4 all have two links linked to them, and we want to compute the emotion value of T3 and T4. According to intensity assumption, T4 has stronger emotion than T3 in emotion ek.

Not all of the words in WEAN of ek have this emotion. Actually, the ek of many words is almost 0, which can be ignored. These words in WEAN are mainly those whose degrees are small. So, when computing word emotion, we should set these words emotion of ek null, and the other word emotion value is between 0 and 1. The number of words whose emotion should be set null can be decided automatically.

Based on Quantity Assumption, Intensity Assumption and the weights of the links in Definition 1, the emotion of word kn can be computed as,

u kekn

(j) = N∑ i=1

V kekn

(j−1)×ωe k

i,n (3)

and

V kekn

(j) = f ( u kekn

(j)/ε )

(4)

where V kekn

(j−1) denotes the (j−1)-th computation value of

word kn on ek. Word emotion value plus link weight reflects the intensity assumption. It means that word ke

k

i transmits its emotion to word ke

k

n . The summation reflects quantity assumption. u

kekn (j) is the storage that stores the sum of trans-

mitting values. f (x) = 1−e −x

1+e−x is the normalized function, who normalizes the words’ value between 0 and 1. ε is a regulation parameter that regulates most word emotion to 0.

When computing word emotion, we randomly give each word an initial value. The computation can yield rational results only after finite times of iterations. Eq. (3) and (4) is for a single word. For all of the words in WEAN, Eq. (5) and (6) work.( u ke k

1 (j), · · · ,u

ke k N (j) )

=

( V ke k

1 (j−1), · · · ,V

ke k N (j−1)

) ·

  ω

ek 1,1 · · · ω

ek 1,N

... ...

...

ωe k

N,1 · · · ω ek N,N

  (5)

and

( V ke k

1 (j), · · · ,V

ke k N (j) ) = f

  ( u ke k

1 (j), · · · ,u

ke k N (j) )

ε

  (6)

From the above procedure, it can be seen that the emo- tion of a word is determined by the emotion of the words who have links with it. The link is like to vote this word.

The emotion of a word is obtained by a recursive algorithm for all of the words linked with it. A word who has more links and the emotion value of linked words will tend to have a larger emotion value. The whole procedure is summarized in Algorithm 1.

Algorithm 1 Word Emotion Computation Require: WEAN and wei,j Ensure: Word emotions 1: randomly initialization 2: tag = 1, it=1; 3: for each kth dimension of emotion do 4: for tag = 1 do 5: tag = 0; 6: for each word wi in WEAN do 7: Update Ve

k

wi (it) by Eq. (5) and (6); 8: end for 9: for each word wi in WEAN do

10: if |Ve k

wi (it)−V ek wi (it −1)|> 0 then

11: tag = 1; 12: end if 13: end for 14: end for 15: end for

B. WORD EMOTION REFINEMENT THROUGH STANDARD SENTIMENT THESAURUS The proposed algorithm for word emotion computation above only considers the emotion of emoticons in microblogs. How- ever, in the early stage of one event, microblogs about this event is not many, especially the microblogs with emoticons. The WEAN constructed by few words and rules is lack of credibility. At this point, the accuracy of using WEAN to compute the word emotion will be low. While the words in standard sentiment thesaurus have obvious and stable emo- tion, which seldom change with the different news events. Therefore, we plan to refine the word emotion obtained by Algorithm 1 according to the words in standard sentiment thesaurus.

For the words in both standard sentiment thesaurus of ek

and WEAN, we define them as wB. Emotion ek of these words is approximate to 1, actually. However, the words emotion computed by Algorithm 1 may be less than 1. At this time, we need to refine them. We set the maximum error for words in wB is δ, and every step is 1, which is a positive number. When the emotion of one word in wB computed by iteration is much less than 1 (i.e., |1−Ve

k

wb| > δ), we add the weights of

the links with the word 1·|1−Ve

k wb |

δ . After that, we do the word

emotion computation again using Algorithm 1. When all the words (in wB) emotion values are close to 1, we stop the word emotion computation. The whole procedure is summarized in Algorithm 2.

Texts are composed by words. Words’ emotion reflect texts’ emotion indirectly. After computing word emotion, we

VOLUME 5, 2017 2377

D. Jiang et al.: Sentiment Computing for the News Event Based on the Social Media Big Data

Algorithm 2 Word Emotion Refinement Require: WEAN , weij, δ, 1 and standard sentiment

thesaurus Ensure: Word emotions 1: Word emotions obtained by Algorithm 1; 2: tag = 1; 3: for each k − th dimension of emotion do 4: for tag = 1 do 5: tag = 0; 6: for each word i both in standard sentiment the-

saurus and WEAN do 7: if |1−Ve

k

wb| > δ then 8: tag = 1; 9: for each link weight we

k

i,j linked with word i do

10: we k

i,j = w ek i,j +1 · frac|1−V

ek wb|δ

11: end for 12: end if 13: end for 14: if tag = 1 then 15: for each word wi in WEAN do 16: Update Ve

k

wi by Eq. (5) and (6); 17: end for 18: end if 19: end for 20: end for

can obtain text six-dimensional emotion by adding the six- dimensional emotion of words that in the text, respectively.

es = ∑ i:wi∈S

ewi (7)

where ewi =< e love wi ,e

joy wi ,e

anger wi ,e

sad wi ,e

fear wi ,e

surprise wi > is

a six-dimensional vector, which means the six-dimensional emotion of wi; es is also a six-dimensional vector, which means the six-dimensional emotion of text.

IV. EXPERIMENTS In this section, we conduct some experiments to validate the correctness and effectiveness of the proposed method.

A. DATASETS To evaluate our method, we use the news event The Malaysia Airlines MH370 on March 8, 2014 as dataset. All of the social media data about this news event are from Sina Microblog. It includes 48,396 microblogs from March 8, 2014 to April 6, 2014. Among the thirty days’ texts, 7,346 of them have emoticons, and these can be used to compute words emotion.

B. EVALUATION METRIC In order to evaluate performance of the proposed method on the text affective computing, we need to design an evaluation metric first. Our idea is that the emotion of a microblog should be consistent with its emoticons since microblogs are

very short. Therefore, the emotion computed from emoticons could be seen as the benchmarks of the proposed method as followings,

edb = 1 Nde

Nde∑ i=1

eei (8)

where Nde is the number of emoticons in microblog d; ei is one emoticon in microblog d; eei is the emotion vector of an emoticon; edb is the emotion vector of microblog d. e

d b could

be seen as the benchmark of our proposed method, so we can evaluate the performance of the proposed method through comparing edb with the value from our method.

C. EXPERIMENTAL SETUP Since both the proposed method and the evaluation metric need the microblogs with emoticons, we need to split the data into two parts: one part as training data and the other as test data. The whole procedure is as follows:

1) Pick the microblogs with emoticons of a news event; 2) Split these microblogs with emoticons into two parts:

training part and test part; 3) Put the training part with the microblogs wiping off

emoticons into our proposed method; 4) Predict the test part using the the proposed well-trained

method; 5) Compute the benchmarks for test part using Eq. 8; 6) Evaluate the performance of the proposed method.

D. EXPERIMENTAL RESULTS AND DISCUSSIONS The word emotion computation, as shown in Algorithm 1, needs the initialization of the variables. A nature question may be asked: is this algorithm sensitive to the different initializations? The answer is no. Here, we are going to empirically show that Algorithm 1 is not sensitive to the different initializations. We feed the Algorithm 1 different initializations and run it several times. The values at each iteration are recorded and plotted in Fig. 5. It can be seen from Fig. 5 that the different starting points will lead to same con- vergent value for the same word after a number of iterations (around 9 in this example). This conclusion improves our confidence about the output of Algorithm 1 even with random initialization. Furthermore, we plotted the convergent curves of ten different words in Fig. 6. Not surprisingly, the emotion values of all the words are convergent after a number of iter- ations. The more interesting observation is that the different words will converge to different values even with different or same initializations. Therefore, we can draw the conclusion that the final emotion values for words from Algorithm 1 are not influenced by the initialization but determined by their structural portions in the WEAN.

In order to show the effectiveness of word emotion refine- ment in Algorithm 2, we conduct eight group experiments to compare the word emotions before and after the refine- ment. From Table 2, all eight groups have reached an accu- racy of more than 75%. At the same time, the line chart

2378 VOLUME 5, 2017

D. Jiang et al.: Sentiment Computing for the News Event Based on the Social Media Big Data

FIGURE 5. The changing of word emotions for one same word with different initializations during the iteration.

FIGURE 6. The changing of word emotions for ten different words during the iteration.

TABLE 2. The accuracy of text emotion computation.

in Fig. 7 shows that the refined emotion is more accurate than before,but not very obvious. The reason for that is we acquire all microblogs about The Malaysia Airlines MH370 and the WEAN based on that is totally integrated. So, we can obtain a result even more accuracy than the refined one.

In Fig. 8, we plot word emotions of 5,363 words from news event The Malaysia Airlines MH370. Since each word

FIGURE 7. The accuracy of text emotion computation.

emotion is a 6-dimensional vector, there are six subfiguers in Fig. 8 each of which corresponds to one dimension of word emotion (red one is for love; pink one is for joy; green one is for anger; blue one is for sad; black one is for fear; cyan one is for surprise). Each point in the subfiguer denotes a word and its corresponding value in y-axis is its final emotion value on this dimension. It can be seen from Fig. 8 that the emotion love and sad of these words are much more stronger than the other four dimensions of emotion because there are more words located in [0.51] in subfiguers love and sad. So, we can draw the conclusion that people mainly feel love and sad on event The Malaysia Airlines MH370. In order to show the accuracy of our method, we have

compared our method that sums word emotions together as the text emotion with benchmark. In Table 3, we have shown six microblog (see Fig. 9 from TheMalaysiaAirlinesMH370) examples with their emotion from both our method and the benchmark emotion using developed evaluation metric in Eq. 8. The result shows that although the short text is not with single emotion, the main dimension’s emotion is in keeping with the benchmarks. Therefore, we can draw the conclusion that our method is effective in short text emotion computation for the news event.

We define the summation text emotion of all microblogs of a news event as the social emotion of this news event at a given time. Fig. 10 shows the emotion of The Malaysia Airlines MH370 from 2014.03.08 to 2014.04.06. The change/evolving of the emotion is due to the develop- ment of this event and some new information has joint the public discussions. It can be found that emotion of love and sad has the similar trend. The x-axis represents the dates. 2014.03.08 was the date that MH370 flight lost contact and then Malaysia Airlines proved this information. The Chinese government launched the emergency mechanism. In the next few days, people are increasingly concerned about MH370 flight and all hope to find the airplane. On 2014.03.13, the suspect of the military covering the true information about the

VOLUME 5, 2017 2379

D. Jiang et al.: Sentiment Computing for the News Event Based on the Social Media Big Data

FIGURE 8. Visualization of the word emotions for 5,363 words of news event The Malaysia Airlines MH370.

TABLE 3. Text emotion for the six microblogs in Fig. 9.

FIGURE 9. Six microblogs with emoticons.

event leading to people’s peak attention. Several days after that, people lose their concentration due to the slow search and rescue speed. On 2014.03.24, Malaysia Prime Minister declares that the airplane dropped in south India Ocean and no one survived. This information led people’s attention back to MH370 incident in few days. On 2014.03.27, the Malaysia

FIGURE 10. Evolution of the emotion of The Malaysia Airlines MH370.

government confirmed 122 pieces of debris found in the waters west of Perth and Australian Maritime Bureau said the search staff had witnessed 3 objects. On 2014.04.02, Malaysia Airlines MH370 survey was defined as a criminal investigation. In this process, the development of the incident has caused the change of social emotion, and the public opinion tends to be stable with the evolution of the event.

2380 VOLUME 5, 2017

D. Jiang et al.: Sentiment Computing for the News Event Based on the Social Media Big Data

V. CONCLUSIONS AND FUTURE STUDY In this paper, we have developed an innovative method to do the sentiment computing for the news events based on the social media big data. The proposed method consists two procedures: word emotion computation through word emotion association network and word emotion refinement through standard emotion thesaurus. For the word emotion computation through word emotion association network, a Word Emotion Association Network (WEAN) has been built to jointly express its semantic and emotion, which is the basis for both word and text emotion computations. Based on WEAN, a word emotion computation algorithm has been proposed to obtain the initial word emotions through a designed iterative process which convergence has also been proofed. Furthermore, a word emotion refinement algorithm has been proposed to improve the accuracy by incorporating the common prior knowledge: standard emotion thesaurus. The experiments have also demonstrated the necessary of this refinement. At last, we choose The Malaysia Airlines MH370 event to evaluate our method’s effectiveness. The result shows that our method can compute news event sentiment exactly.

In the future, we are interested in considering emotion distance and word emotion pattern into text sentiment com- putation. One microblog is published by one person and with only less than 140 words. It should only have single strong emotion. And one word in different circumstances my have different emotion. So when computing text emotion, we should consider word context. Furthermore, Brain-like Computing also needs to compute sentiment. So our work can be used into Brain-like Computing using transfer learning method.

ACKNOWLEDGMENTS The research work reported in this paper was supported in part by the Shanghai Science International Cooperation Project under grant no. 16550720400. This work was jointly sup- ported by the National Science Foundation of China under grant no. 61471232 and the innovation project of Institute of Computing Technology (ICT), Chinese Academy of Sci- ence (CAS).

REFERENCES [1] S. Zhang, J. Liang, R. He, and Z. Sun, ‘‘Code consistent hashing based

on information-theoretic criterion,’’ IEEE Trans. Big Data, vol. 1, no. 3, pp. 84–94, Sep. 2015.

[2] J. Xuan et al., ‘‘Uncertainty analysis for the keyword system of web events,’’ IEEE Trans. Syst., Man, Cybern., Syst., vol. 46, no. 6, pp. 829–842, 2016.

[3] W. McDougall, An Introduction to Social Psychology. North Chelmsford, MA, USA: Courier Corporation, 2003.

[4] B. Pang, L. Lee, and S. Vaithyanathan, ‘‘Thumbs up?: Sentiment classifi- cation using machine learning techniques,’’ in Proc. ACL Conf. Empirical Methods Natural Lang. Process., 2002, pp. 79–86.

[5] H. Cui, V. Mittal, and M. Datar, ‘‘Comparative experiments on sentiment classification for online product reviews,’’ in Proc. AAAI, vol. 6. 2006, pp. 1265–1270.

[6] S.-M. Kim and E. Hovy, ‘‘Automatic identification of pro and con reasons in online reviews,’’ in Proc. COLING/ACL Main Conf. Poster Sessions, 2006, pp. 483–490.

[7] M. Taboada, J. Brooke, M. Tofiloski, K. Voll, and M. Stede, ‘‘Lexicon- based methods for sentiment analysis,’’ Comput. Linguistics, vol. 37, no. 2, pp. 267–307, 2011.

[8] A. Aue and M. Gamon, ‘‘Customizing sentiment classifiers to new domains: A case study,’’ in Proc. Recent Adv. Natural Lang. Process. (RANLP), 2005, vol. 1. nos. 3–1, pp. 1–2.

[9] S. Blair-Goldensohn, K. Hannan, R. McDonald, T. Neylon, G. A. Reis, and J. Reynar, ‘‘Building a sentiment summarizer for local service reviews,’’ in Proc.WWWWorkshopNLPInf.ExplosionEra, vol. 14. 2008, pp. 339–348.

[10] S. Bao et al., ‘‘Mining social emotions from affective text,’’ IEEE Trans. Knowl. Data Eng., vol. 24, no. 9, pp. 1658–1670, Sep. 2012.

[11] Y. Rao, Q. Li, X. Mao, and L. Wenyin, ‘‘Sentiment topic models for social emotion mining,’’ Inf. Sci., vol. 266, pp. 90–100, May 2014.

[12] Y. Yamamoto, T. Kumamoto, and A. Nadamoto, ‘‘Multidimensional sen- timent calculation method for twitter based on emoticons,’’ Int. J. Pervas. Comput. Commun., vol. 11, no. 2, pp. 212–232, 2015.

[13] D. Tang, B. Qin, and T. Liu, ‘‘Learning semantic representations of users and products for document level sentiment classification,’’in Proc.Meeting Assoc. Comput. Linguistics Int. Joint Conf. Natural Lang., vol. 1. 2015, pp. 1014–1023.

[14] J. Read, ‘‘Using emoticons to reduce dependency in machine learning tech- niques for sentiment classification,’’ in Proc. ACL Student Res. Workshop, 2005, pp. 43–48.

[15] A. Go, R. Bhayani, and L. Huang, ‘‘Twitter sentiment classification using distant supervision,’’ Stanford Univ., Stanford, CA, USA, Project Rep. CS224N, 2009, p. 12.

[16] A. Pak and P. Paroubek, ‘‘Twitter as a corpus for sentiment analysis and opinion mining,’’ in Proc. LREc, vol. 10. 2010, pp. 1320–1326.

[17] L. Zhang, R. Ghosh, M. Dekhil, M. Hsu, and B. Liu, ‘‘Combining lexicon- based and learning-based methods for twitter sentiment analysis,’’ HP Lab., Palo Alto, CA, USA, Tech. Rep. HPL-2011-89, 2011.

[18] D. Davidov, O. Tsur, and A. Rappoport, ‘‘Enhanced sentiment learning using twitter hashtags and smileys,’’ in Proc. 23rd Int. Conf. Comput. Linguistics, Posters, 2010, pp. 241–249.

[19] A. Bakliwal, P. Arora, S. Madhappan, N. Kapre, M. Singh, and V. Varma, ‘‘Mining sentiments from Tweets,’’ in Proc. WASSA, 2012, pp. 11–18.

[20] L. Jiang, M. Yu, M. Zhou, X. Liu, and T. Zhao, ‘‘Target-dependent twitter sentiment classification,’’ in Proc. 49th Annu. Meeting Assoc. Comput. Linguistics, Human Lang. Technol., 2011, pp. 151–160.

[21] X. Hu, J. Tang, H. Gao, and H. Liu, ‘‘Unsupervised sentiment analysis with emotional signals,’’ in Proc. 22nd Int. Conf. World Wide Web, 2013, pp. 607–618.

[22] H. Saif, Y. He, and H. Alani, Semantic Sentiment Analysis of Twitter. Berlin, Germany: Springer, 2012.

[23] H. Saif, Y. He, H. Alani, and M. Fernandez, ‘‘On stopwords, filtering and data sparsity for sentiment analysis of Twitter,’’ in Proc. Int. Conf. Lang. Resour. Eval., 2014, pp. 810–817.

[24] H. Saif, Y. He, M. Fernandez, and H. Alani, ‘‘Contextual semantics for sentiment analysis of Twitter,’’ Inf. Process. Manage., vol. 52, no. 1, pp. 5–19, 2015.

[25] X. Luo, Z. Xu, J. Yu, and X. Chen, ‘‘Building association link network for semantic link on Web resources,’’ IEEE Trans. Autom. Sci. Eng., vol. 8, no. 3, pp. 482–494, Jul. 2011.

DANDAN JIANG received the bachelor’s degree from Shanghai University, China, in 2014, where she is currently pursuing the master’s degree. Her main research interests include social media text sentiment computation and machine learning.

VOLUME 5, 2017 2381

D. Jiang et al.: Sentiment Computing for the News Event Based on the Social Media Big Data

XIANGFENG LUO received the master’s and Ph.D. degrees from the Hefei University of Tech- nology in 2000 and 2003, respectively. He was a Post-Doctoral Researcher with the China Knowl- edge Grid Research Group, Institute of Com- puting Technology, Chinese Academy of Sci- ences, from 2003 to 2005. He is currently a Professor with the School of Computers, Shanghai University, China. He is also a Visiting Professor with Purdue University. He has authored or coau-

thored over 50 publications and his publications have appeared in the IEEE TRANSACTION ON AUTOMATION SCIENCE and ENGINEERING, the IEEE TRANSACTION ON SYSTEMS, MAN, and CYBERNETICS-PART C, and the IEEE TRANSACTION ON LEARNING TECHNOLOGY, CONCURRENCY AND COMPUTATION: PRACTICE AND EXPERIENCE, AND NEW GENERATION COMPUTING. His main research interests include Web wisdom, cognitive informatics, and text understanding. He has served as the Guest Editor of the ACM Transactions on Intelligent Systems and Technology. He has also served on the committees of a number of conferences/workshops, including the Program Co-chair of ICWL 2010 (Shanghai), WISM 2012 (Chengdu), CTUW2011 (Sydney), and more than 40 PC members of conferences and workshops.

JUNYU XUAN received the bachelor’s degree from the China University of Geosciences, Beijing, in 2008. He is currently pursuing the Dual Ph.D. degree with Shanghai University and the University of Technology Sydney. His main research interests include machine learning, com- plex network, and Web mining.

ZHENG XU was born in Shanghai, China. He received the Diploma and Ph.D. degrees from the School of Computing Engineering and Science, Shanghai University, Shanghai, in 2007 and 2012, respectively. He is in a post- doctoral position with Tsinghua University from 2014 to 2016. He is currently with the third research institute of the Ministry of Public Secu- rity. His current research interests include mobile crowd sensing, public security, and emergency

management. He has authored or coauthored over 80 publications, including IEEE TRANSACTION ON FUZZY SYSTEMS, the IEEE TRANSACTION ON AUTOMATION SCIENCE AND ENGINEERING, the IEEE TRANSACTION ON CLOUD COMPUTING, the IEEE TRANSACTION ON EMERGING TOPICS IN COMPUTING, and the IEEE TRANSACTION ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS.

2382 VOLUME 5, 2017