Emotion Extraction and Entrainment in Social Media: The Case of U.S. Immigration and Border Security

Transcription

1 Emotion Extraction and Entrainment in Social Media: The Case of U.S. Immigration and Border Security Wingyan Chung Dept. of Decision and Info. Sciences, School of Business Administration, Stetson University DeLand, Florida, U.S.A. Saike He State Key Lab of Management and Control for Complex Systems, Chinese Academy of Sciences Beijing, China Daniel Dajun Zeng Dept. of MIS, U. of Arizona Tucson, Arizona, U.S.A. and Chinese Academy of Sciences Beijing, China Victor Benjamin Dept. of MIS Eller College of Management University of Arizona Tucson, Arizona, U.S.A. Abstract Emotion plays an important role in shaping public policy and business decisions. The growth of social media has allowed people to express their emotion publicly in an unprecedented manner. Textual content and user linkages fostered by social media networks can be used to examine emotion types, intensity, and contagion. However, research into how emotion evolves and entrains in social media that influence security issues is scarce. In this research, we developed an approach to analyzing emotion expressed in political social media. We compared two methods of emotion analysis to identify influential users and to trace their contagion effects on public emotion, and report preliminary findings of analyzing the emotion of 105,304 users who posted 189,012 tweets on the U.S. immigration and border security issues in November The results provide strong implication for understanding social actions and for collecting social intelligence for security informatics. This research should contribute to helping decision makers and security personnel to use public emotion effectively to develop appropriate strategies. Keywords emotion; social media analytics; text mining; social network analysis; entrainment; influence; U.S. immigration policy; border security. I. INTRODUCTION Emotion plays an important role in shaping public policy and business decisions. Politicians and marketers often create their agendas based on public emotion expressed in mass media. As social media are increasingly used to share ideas and opinions, analyzing the intensity and types of emotion in these messages can help decision makers develop more appropriate strategies to address the needs of their constituents. The United States (U.S.) immigration policy has recently gained widespread attention on social media as well as on national and international media. Because the U.S. population consists primarily of immigrants or descendants of immigrants, public policy on immigration has strategic and long-term effects on the economy, foreign policy, and national security [1]. As nearly twelve million undocumented immigrants reside in the United States, the U.S. Government and Congress are under tremendous pressure of developing timely legislations for reforming immigration policies, which have a strong impact on U.S. border security. How to gauge the emotion of the public on border security, immigration, and related policy issues becomes a concern for decision makers and politicians. Unfortunately, there is little work on social-media-based security informatics that could potentially benefit related policy decision making. By analyzing the content and linkage information in social media, policy makers and social scientists interested in immigration and border security could possibly obtain useful answers to such questions as: Who are the most influential leaders in the social media community? How are different types of emotion changing over time as new events emerge? How does emotion spread and entrain in the community? What are the network relationships among participants in immigration and border security discussion on social media? In this research, we developed an approach to collecting and analyzing emotion expressed in political social media. We compared two methods of emotion analysis to identify influential users and to trace their contagion effects on public emotion, and report preliminary findings of analyzing the emotion of 105,304 users who posted 189,012 tweets on the U.S. immigration and border security issues during November The results provide strong implication for understanding social actions and for collecting social intelligence for security informatics. II. LITERATURE REVIEW The analysis of emotion in social media relates to literature in social computing, social media and network analysis, and emotion entrainment. We review below these domains with a view of informing the analysis of emotion in social-mediabased security informatics. A. Social Media Analytics in Security Informatics Social media analytics (SMA) is a rich set of tools and technologies most recently applied to security informatics [2, 3]. Traditional domains that SMA is used include business intelligence, online product review analysis, and security informatics [4-6]. Among these technologies, sentiment analysis and emotion extraction are important tools for understanding large amounts of social media data [7]. Researchers have developed a specialized lexicon to support emotion extraction [8]. The lexicon contains over 13,000 1

2 entries of term-category, in which a term may belong to one or more of the eight emotion categories (anger, anticipation, disgust, fear, joy, sadness, surprise, trust) [9]. Network analysis is another SMA application in which users within a large community are analyzed to reveal their levels of influence. Traditionally, social network analysis examines network dynamics in a small scale [10]. Advances in information technologies enable new capabilities of network analysis, such as identifying opinion leaders in discussion forums [11], finding expertise in networks [12], analyzing business stakeholder networks [13], and visualizing networked relationships of infectious disease transmission [14]. Although social media are increasingly important information sources for understanding public sentiment, little work is found in analyzing emotion to support intelligence and security informatics decision making. B. Emotion Analysis and Entrainment The proliferation of social media has made them the hub of information generation and diffusion. One important problem is to characterize and identify influentials, which can be defined as users who influence the opinions or emotions of others in large scale. Previous work quantifies user influence based on various structural measures, such as in/out-degree, betweenness, and PageRank [15]. The idea is based on social linkages developed in human interaction that facilitate spread of human emotion. These linkages play a critical role in large social media communities by facilitating exchange of information and confirmation of authoritative and influential positions. However, they have not been examined widely in social-media-based security informatics. Recent works also attempt to model the spread of human behavior in social networks [16-19]. Emotion entrainment, which accounts for the synchronous convergence of human emotions, has been examined in neuroscience [20] and in human social networks [21]. Whether human emotion entrains in large-scale networks of importance to national security and public policy decision making is relatively unknown. While previous methods provide quantitative analysis for social interactions, they often require explicit casual knowledge that is scarce in many scenarios. III. APPROACHES TO EMOTION ANALYSIS FOR SOCIAL- MEDIA-BASED SECURITY INFORMATICS In this research, we developed an approach to emotion analysis for social-media-based security informatics. Our approach consists of emotion extraction and entrainment of a large-scale social media network of significance to national security and public policy decision making. Through applying our approach, we examined the process of identifying influential leaders in social media networks, studied changes in emotion over time, and revealed network relationships among participants. We describe below our dataset, analyses, and preliminary findings of the two approaches to emotion analysis. A. Dataset We collected the tweets that are related to the U.S. immigration and border security discussion. To do so, we constructed carefully a list of queries by reviewing recent literature published on the topic [1, 22-24]. Filtering and collection testing helped to reduce the original 14 queries to 8 domain-specific queries: comprehensive immigration reform, illegal alien, illegal immigration, immigration debate, immigration policy, immigration reform, US border security, and US immigration. We developed a tweet crawler that collects tweet postings automatically and periodically since May As of February 15, 2015, our dataset consists of 2,418,590 tweets posted by 617,536 unique users. In this case study, we focus on the recent executive actions announced by President Obama in November 2014 to grant legal status to millions of illegal immigrants [25, 26]. We selected a subset of the data consisting of 189,012 tweets posted by 105,304 users between November 1st and November 30th of That month also recorded the highest number of tweets and the most user activities during the entire timespan of the data collection. From the users whose tweets were collected in November, we randomly selected a sample of 5,000 users for further examination of the emotion extracted from the tweets and of emotion entrainment. B. Emotion Extraction To extract emotion from each tweet, we computed values of indices in eight emotional categories defined in [9] (anger, disgust, fear, sadness, surprise, anticipation, joy, and trust). Each index represents the emotional intensity of a tweet. We used the emotion lexicon developed in [8] to identify emotional words that appear in the tweets. The lexicon contains 13,901 word-emotion entries tagged by over 2,216 human users who completed a total of 38,726 tagging assignments. Each word was assigned a sentiment polarity (positive (+1) or negative (- 1)) and is categorized into one (or more) of the eight emotional categories. Equation (1) shows the computation of the index of emotion j (ranging from 1 to 8) of tweet i, which contains n i words. M ij = where Pn i W k i=1 n i 8i 2 {1...N}, 8j 2 {1...k}...(1) M ij = Index of emotion j of Tweet i n i = Number of words in Tweet i k = Number of di erent emotion categories N = Total number of tweets being considered 8 < 1 if word(k) is in category j, where W k = k 2 {1...n i } : 0 otherwise C. Emotion Entrainment Emotion entrainment is the socialization of feelings among people who interact with each other. To examine how this process takes place in a large-scale social media network and how emotion evolves dynamically, we developed a model-free approach to detect causal relationships and to identify influential users based on the theory of emotion entrainment [27]. This theory suggests that human emotion rhythmically converges through social interactions. In this process, we consider users that are most entrained by others as influential. 2

3 We represent the emotion time series of two users x and y with two Markov processes X = x t and Y = y t. Then the entrainment strength from x to y is given by Equation (2). m m n ( X Y ) = ( t+ 1 t ) ( t+ 1 t, t ) En H x x H x x y (2) m n where, x = ( x x ), y = ( y y ), while m and n t t,..., t m+1 t t,..., t n+1 are the orders of each of the Markov processes; H(*) calculates the entropy of the probability distribution enclosed. In our experiments, we set the Markov orders m=n=3. According to our empirical analysis, this order is high enough while higher order increases computational cost without present much quantitative differences. Equation (2) can be estimated based on Simpson s rule [28] with a computational complexity of O(N log(n)). This cost is acceptable for entrainment analysis on large-scale datasets. Using Equation (3), we measured the influence of each user x based on a how others entrain towards him. ( ) = ( ) I x En y x (3) x y IV. PRELIMINARY FINDINGS In this section, we present preliminary findings of using the emotion analysis methods to examine influence of Twitter TABLE I. Social-Network-Based Method users who posted on immigration and border security issues in November 2014, and to provide descriptive statistics of the emotion scores. A. Comparison of Top Influential Users TABLE I. lists the top 10 most influential users identified by two methods. The first method uses the Betweenness Centrality score as the measure of influence of a user in the social media network formed from connections of retweets and modified-tweets. While other measures of centrality are available [15, 29], Betweenness Centrality (BC) identifies the extent to which a user serves as a bridge in immigration reform discussion, which heavily relies on the ability to find compromises and common grounds in participants of highly polarized views. The resulting top users are shown on the left of the table. Several of the listed users are out-spoken leaders in U.S. immigration and border security issues. For example, Daniel John Sobieski identified himself as a free-lance writer for various news agencies and expresses strong opinions on the immigration policy proposed by President Obama. FAIR (Federation for American Immigration Reform) is concerned with border security, immigration law enforcement, amnesty, and stopping illegal immigration. TOP 10 MOST INFLUENTIAL USERS IDENTIFIED BY TWO METHODS Emotion Entrainment Method User Name URL BC Score User Name URL Entrain Score Daniel John Sobieski ladybug FAIR SHEEPDOG REPORT Michael Johns CONEXION Mickey Kaus Marty Barrack daveweigel Stellasasha? Greg Sargent Les Wenzel Jeremy W. Peters Mike Robbins Matthew Kolken Ted Skibinski Julia Preston Ken Michael Deacon Dave Ganteaume The second method uses Equation (3) as a measure of a user s influence in a social media network. In contrast with the results found by the first method, among the top three users found by the entrainment method, only CONEXION lists itself as related to Latino entrepreneurs, which may be related to immigration issues because many immigrants are Latinos. B. Distribution of emotion ratio To understand the relative strength of the emotion, we computed the ratio for the j th emotion e in top k users (sorted j by descending order of influence score calculated by (3)) as: i=1 ( ) = k k m R e j k e ij i=1 j=1 e ij, j =1,!,m; k =1,!,n where, ij e is the sum of all the j th emotion scores (among m (= 8) categories defined in [9]) of the tweets written by the i th user in the ranked list of n (= 5,000) users. Intuitively, R(e j ) k characterizes the emotion distribution over the top k users. If R(e j ) k varies significantly as k increases (as more less-influential users are involved in the calculation of ratio), (4) 3

4 the j th emotion should be considered as a critical factor for distinguishing influential users from normal users. while surprise was the least dominant emotion (with the lowest ratio). Third, the ratio for trust increases dramatically as users ranks change from 1 to n (= 5000). This means the relative amount of trust expressed in social media increasingly distinguishes influential users from less influential users, i.e. expressing less trust in social media increases user influence in the community. C. Entrainment network In this section, we examine the structure of the entrainment patterns that affect the influence of each user. The entrainment networks of twitter users named Daniel John Sobieski, SHEEPDOG REPORT and ANNA RAND are shown in Figures 2, 3(a) and 3(b) respectively. All nodes are labeled by the names of users who entrain toward the specified user (central node labeled yellow). Blue nodes have above-average entrainment strengths while white nodes have below-average strengths. The arrow on an edge indicates the entrainment direction and the thickness of an edge represents entrainment strength. We first explore two users who had the highest BC score (see Table I). The ratios of users entraining towards these two users with above-average strengths are respectively 41.53% and 42.86%. These similarly high ratios suggest that influential users are likely to be entrained by other others with similar degree of entrainment. To test this hypothesis, we further compared the second most influential user SHEEPDOG REPORT found by the second method (see Figure 1. Relative strengths of eight emotion types. Figure 1 illustrates the change in ratio of emotions in top n (= 5000) most influential users. There are three aspects worth mentioning. First, the ratios of fear, disgust, anger, and sadness decrease while the ratios of trust, joy, anticipation, and surprise increases. These results indicate that influential users are more likely to be those high in emotions related to fear, disgust, anger, and sadness. Second, fear was the most dominant emotion (with the highest ratio) in the posted tweets Figure 2. Entrainment network of Daniel John Sobieski. 4

5 security issues during November Our approach uses a comprehensive lexicon to identify emotion categories, includes algorithms to compute emotion scores and user influence scores (based on both social ties and entrainment), and present analysis findings of users network relationship and influence patterns. These results should provide strong implication for understanding social actions and for collecting social intelligence for security informatics. Table I) and the second least influential user ANNA RAND (URL: (found by the second method but not listed in Table I). These two users tweeted similar numbers of messages (46.6K v.s. 52.1K), yet had totally different influence in discussing immigration and border security in the community. The ratios of users with above-average entrainment strengths are 57.93% for SHEEPDOG REPORT and 32.00% for ANNA RAND (note the relatively sparse edges compared with the other networks shown). This result supports our hypothesis that entrainment strengths among influential users are similar to each other. The findings provide strong implications for decision makers to identify influential users or leader groups in a social media community, and for users who aspire to become leaders in their communities. This research has several limitations. First, the generalpurpose emotion lexicon used in this research was not designed specifically for the our domain and for short messages such as tweets. Future study may verify the effectiveness of such lexicon different domains and different types of messages. Second, only Twitter was used to provide social media messages, despite other SM platforms may provide abundant content in the domain. Third, our data collection was based on pre-defined keyword queries that are restricted in terms of coverage. V. SUMMARY The growth of social media enables users to express their emotion publicly and promptly, thereby helping politicians and business leaders to gauge public perception of their agendas. Textual content and user linkages fostered by social media networks can be useful in examining emotion in a large-scale manner. However, research into how emotion evolves and entrains in social media that influence security issues is scarce. In this research, we developed an approach to analyzing emotion expressed in political social media. We compared two methods of emotion analysis to identify influential users and to trace their contagion effects on public emotion, and report preliminary findings of analyzing the emotion of 105,304 users who posted 189,012 tweets on the U.S. immigration and border Our ongoing works include conducting a thorough evaluation of the emotion analysis results, developing metrics to understand performance, and testing the approach on new datasets. This research should contribute to helping decision makers and security personnel to use public emotion effectively to develop appropriate strategies. ACKNOWLEDGEMENTS This paper is based upon work supported partially by funding from the U.S. National Science Foundation (through project DUE ), U.S. Department of Homeland Security (through the 2013 Summer Research Team Program (a) SHEEPDOG REPORT (b) ANNA RAND Figure 3(a) and (b). Entrainment networks of (a) SHEEPDOG REPORT and (b) ANNA RAND. 5

6 for MSI), by Stetson University s 2015 Summer Research Grant, by the National Natural Science Foundation of China under Grant No , , , , and , by the U.S. National Institute of Health under Grant No. 1R01DA , by the China Ministry of Health under Grant No. 2013ZX Any opinions, findings, and conclusions or recommendations expressed in this paper are those of the authors and do not necessarily reflect the views of the funding agencies. We thank Mr. Daniel Ballard, the conference chairs, reviewers, and coordinators for their assistance and valuable suggestions. REFERENCES [1] J. Bush, T. F. McLarty III and E. H. Alden, U.S. immigration policy. New York: Council on Foreign Relations, [2] D. Zeng, H. Chen, R. Lusch and S.-H. Li, "Social Media Analytics and Intelligence," IEEE Intelligent Systems, vol. 25, pp , [3] V. Benjamin, W. Chung, A. Abbasi, J. Chuang, C. A. Larson and H. Chen, "Evaluating Text Visualization for Authorship Analysis," Security Informatics, vol. 3, [4] H. Chen, R. Chiang and V. Storey, "Business Intelligence and Analytics," MIS Quarterly, vol. 36, pp , [5] W. Chung and T.-L. Tseng, "Discovering business intelligence from online product reviews: A ruleinduction framework," Expert Systems with Applications, vol. 39, pp , [6] W. Chung, "BizPro: Extracting and Categorizing Business Intelligence Factors from Textual News Articles," International Journal of Information Management, vol. 34, pp , [7] A. Abbasi, H. Chen, S. Thoms and T. Fu, "Affect analysis of Web forums and blogs using correlation ensembles," IEEE Transactions on Knowledge and Data Engineering, vol. 20, pp , [8] S. M. Mohammad and P. D. Turney, "Crowdsourcing a Word-Emotion Association Lexicon," Computational Intelligence, vol. 29, pp , [9] R. Plutchik, "A general psychoevolutionary theory of emotion," Emotion: Theory, research, and experience, vol. 1, pp. 3-33, [10] S. Wasserman and K. Faust, Social Network Analysis: methods and applications: Cambridge University Press, [11] X. Song, Y. Chi, K. Hino and B. Tseng, "Identifying opinion leaders in the blogosphere," Proceedings of the sixteenth ACM conference on Conference on information and knowledge management, [12] J. Zhang, M. S. Ackerman and L. Adamic, "Expertise networks in online communities: structure and algorithms," Proc. of the 16th international conference on World Wide Web, Banff, Alberta, Canada, [13] W. Chung, H. Chen and E. Reid, "Business stakeholder analyzer: An experiment of classifying stakeholders on the Web," Journal of the American Society for Information Science and Technology, vol. 60, pp , [14] D. Zeng, H. Chen, C. Castillo-Chavez, W. B. Lober and M. Thurmond, Infectious Disease Informatics and Biosurveillance, vol. 27: Springer, [15] M. O. Jackson, Social and Economic Networks. Princeton, NJ: Princeton University Press, [16] R. M. Bond, C. J. Fariss, J. J. Jones, A. D. I. Kramer, C. Marlow and J. E. Settle, "A 61-million-person experiment in social influence and political mobilization," Nature, vol. 489, pp , [17] G. C. Kane, M. Alavi, G. J. Labianca and S. P. Borgatti, "What's different about social media networks? A framework and research agenda," MIS Quarterly, vol. 38, pp , [18] Y. Liang, X. Zheng, D. D. Zeng, X. Zhou, S. J. Leischow and W. Chung, "Exploring How the Tobacco Industry Presents and Promotes Itself in Social Media," Journal of Medical Internet Research, vol. 17, [19] T. Schreck and D. Keim, "Visual analysis of social media data," IEEE Computer, vol. 46, pp , [20] J. Bastiaansen, M. Thioux and C. Keysers, "Evidence for mirror systems in emotions," Philosophical Transactions of The Royal Society B, vol. 364, pp , [21] A. D. Kramer, J. E. Guillory and J. T. Hancock, "Experimental evidence of massive-scale emotional contagion through social networks," Proc. of the National Academy of Sciences, vol. 111, pp , [22] J. Gans, E. M. Replogle and D. J. Tichenor, "Debates on U.S. Immigration," Sage Publications, Inc., [23] M. C. LeMay, U.S. immigration: A reference handbook. Santa Barbara, Calif.: ABC-CLIO, [24] U.S. Commission on Immigration Reform, U.S. immigration policy - Restoring credibility: A report to Congress. Washington, DC (1825 Connecticut Ave., NW, Suite 511, Washington 20009): U.S. Commission on Immigration Reform, [25] L. Meckler, C. M. Nelson and E. Morath, "Obama to Protect 4 Million-Plus Immigrants From Deportation," in Wall Street Journal [26] D. Dwyer, "President Obama Offers Legal Status to Millions of Undocumented Immigrants," in ABC News [27] M. Clayton, R. Sager and U. Will, "In time with the music: the concept of entrainment and its significance for ethnomusicology," European Meetings in Ethnomusicology, vol. 11, pp. 1-82, [28] A. Kaw and E. Kalu, Numerical Methods with Applications, 1st. ed: [29] L. C. Freeman, "A set of measures of centrality based on betweenness," Sociometry, vol. 40, pp ,