Polarization Analysis of Twitter Users Using Sentiment Analysis

Transcription

1 Polarization Analysis of Twitter Users Using Sentiment Analysis Nicha Nishikawa, Koichi Yamada, Izumi Suzuki, and Muneyuki Unehara {yamada, suzuki, Department of Information and Management System Engineering, Nagaoka University of Technology, Nagaoka-shi, Japan. Abstract. This work is a preliminary empirical study trying to prove a theory that people who argue on a two-side debate topic are likely to be radical, and to clarify the process of radicalization or polarization through analyzing opinions posted on twitter. Opinions in controversial topics were tracked and analyzed by means of Sentiment analysis, and the results were used to study people's sentiment, for or against, on the matters. Our preliminary results showed that people's opinions were polarized slightly over time in one of the two topics studied, but the other is not. To the best of our knowledge this is the first step to potentially analyze people's opinions in posting about controversial issue. Keywords: Text mining, sentiment analyses, polarization, twitter. 1 Introduction IT technology has been growing dramatically in this century and everyone uses the technology in his/her everyday life, from children and teenagers to adults and elders. In this modern society, the social media play a vital role to our lives. They allow anyone to create and distribute information to his/her friends and family. Social media serve the users as convenient tools of communication, but it is also considered as a powerful gear for social polarization, which seems a serious downside of them. Online news and their comments through the network sometimes cause flooding rudeness and aggression. Twitter, one of mainstream social media, is not an exception, where the unique functions of 140 characters per post and of its global reach have made it become a famous platform for discussion and argument. Twitter has proved to be the voice of people in many situations. Manufacturers can find the feedbacks of their products by collecting customers reviews [1]. People express their mind through tweets not only about the products or services, but for political and debate movement, and it has been intensely studied [2,3]. Since most opinions are available in the text format and its processing is easier than other formats, sentiment analysis has emerged as a subfield of text mining [4,5,6]. Sentiment analysis is an interdisciplinary field that has been used as an opinion recognizer, to judge, to evaluate and to classify the polarity of the given text. It is also 167

2 widely used in different domains such as marketing, politics, entertainment, etc., however as to social issues like controversial topics few research has been done. It is theoretically believed that people who are arguing on a two-side debate topic are likely to be aggressive. Previous studies related to online debate forum have focused on stance classification [7,8], which tried to determine which of the two sides the authors are taking. This paper aims to investigate and gain a better understanding of polarization in debating people and see how changes occur in their emotion as time proceeds. The rest of this paper is organized as follows. Section 2 explains the data and approaches used for this analysis. In section 3, we present the result of our study on the different data topics. Section 4 discusses observations derived from the results. Lastly, we conclude and give future direction of research in section 5. 2 Approach 2.1 Topics of Analysis We have chosen two debate topics for this study. One is Abortion, one of the most controversial topics in the United States, which has been argued hotly so far. The other is Brexit, a widely known political decision of the United Kingdom: the referendum took place in 2016 over a short time period where UK voted to leave or remain in European Union. The final result has become leaving. The difference between these topics is that the resolution has been adopted in the latter but not in the former. For these two debate topics, we collected Twitter data using Hashtag, a type of label starting from "#". We have considered each side as For and Against [7]. In #Brexit, For refers to people who agree with leaving EU, while Against refers to staying. Similarly, For in #Abortion means people who think that abortion should be legal, while Against refer to people who are concerning the embryos' right over women's choice. Twitter offers two APIs to retrieve data, REST API and Streaming API. In this work, in order to analyze changes in emotional mind over time, the historical data was necessary. So the REST API was used. Practically, we used the Twitter Search API, a part of Twitter s REST API to acquire recent tweets with the Hashtags. 2.2 Strategy of Data Collection. The objective of this paper is to analyze the Twitterers' sentiment changes associated with a topic. By using Search API, we selected only the tweets in English that contain the hashtags #Abortion and #Brexit. However, the search API has a limitation, which the search at twitter.com/search does not, that only tweets posted within the past 7 days can be searched. The number of data obtained is, of course, not enough to be analyzed. We performed historical data collection using the strategy described below. First, we identified users who tweeted regularly about the topic in the set of collected tweets, and created a list of the users. Then, each user was used as a parameters to get 168

3 his/her most recent tweets and retweets. This method allows us to get 3,200 tweets from a user's timeline. From the set of tweets, we filtered out only those with the hashtags, #Brexit and #Abortion. Consequently, we could get past tweets from many users who were interested in and discussed the topics. This strategy allowed us to get more historical tweets published before and during the event period. Using the above strategy from July to December 2016, around 100K user accounts who talked about #Brexit and 40K from #Abortion were tracked. The total of collected tweets was 1.5 million tweets, where 340K tweets were from #Abortion which range from 2009 to 2016, and 2.2M tweets were from #Brexit which range from 2013 to Preprocessing Due to the 140 characters limits, Twitter users are so likely to use slang, abbreviation and colloquialisms in their posts. So, before putting them into a classification process, the text preprocessing was applied. In normalization process, all characters were lowercased and repeated characters were replaced with a single (e.g., whyyyy -> why). In addition, because some mentions can be used to improve sentiment classifier, thus, we also converted all username and keyword to a generic tags -> username and #Abortion -> abortion). Finally, we removed additional white space, retweet flag; RT and the presence of URLs. 2.4 Sentiment strength For getting the sentiment score of a tweet, we utilized SentiWordnet, which is a lexicon dictionary developed based on WordNet [9] and is used widely in many studies of the sentiment analysis[10]. It contains WordNet synsets, which are sets of synonyms with the meaning and the part of speech (Noun, Verb, Adjective or Adverb), and are associated with the sentiment degrees of positivity, negativity and objectivity ranging from 0 to 1. The sum of the degrees is 1.0. First, the sentiment score of each word in a tweet is calculated. Given that a word ww has NN senses represented by ww jj where jj is the index of the word sense, the sentiment score of the word ww is given by equation (1): SSSSSSSSSSSSSSSSSS(ww) = NN jj=1 1 jj ssssssssss jj 1 NN jj=1 jj, (1) where ssssssssss jj = PPPPPP jj NNNNNN jj. PPPPPP jj and NNNNNN jj are the positive and negative scores of j-th sense of ww defined by the positive and negative degrees of SentiWordnet, respectively. Then, the sentiment of a tweet is given by the arithmetic average of sentiments of each word contained in the tweet. As discussed previously, FOR in our work refers to people who think positively toward the topic while AGAINST refers to the opposite side. Based on the value of equation (1), we also assigned a label to the numerical value of sentiment strength as 169

4 follows; For AGAINST opinions, we have {strong negative, < -0.5} {negative, to -0.5} {weak negative, 0 to -0.25}, for FOR opinion we have {weak positive, 0 to 0.25} {positive, 0.25 to 0.5} and {strong positive, > 0.5} and {neutral, 0} for the tweet that does not contain any sentiments. The outputs from this process were the sentiment score along with the tweet s published date. These values as well as strength labels were used later to see how change over time. Fig. 1. Average sentiment polarity of tweet for #Abortion Fig. 2. Average sentiment polarity of tweet for #Brexit 170

5 3 Preliminary result The preliminary results can be explained as follows. In order to understand a reflection of public sentiment in each side of each topic, we first calculated the average sentiment score by month of each topic and plotted as a time series. Figure 1 and 2 shows average sentiment polarity from #Abortion and #Brexit respectively. The first part of both topics shows fluctuation of polarity due to the small volume of obtained tweets. For example, in #Abortion of 2009, it amounted only 151 tweets from 16 user accounts. In this number, only 5 tweets were gathered in September 2009 and there was distinctly negative. Similarly, the peaks in #Brexit of 2013 and 2014 responded corresponding to the overall dominant polarity. It could be that Twitter is a new social media and people just getting utilize it to express and discuss their opinions publicly. To visualize the relation between For and Against in each topic, we therefore separated positive tweets from negative tweets. We have noticed that the pulse of the sentiment in a recent years was relatively stable in both #Abortion and #Brexit. As can be seen in Fig. 3 and Fig 4, both #Abortion and #Brexit has average sentiment strength about 0.4 in both positive and negative cases. Fig. 3. #Abortion Sentiment changes over time 171

6 Fig. 4. #Brexit Sentiment changes over time 4 Discussion In this paper, our focus was to study the changes in sentiment of people who gave opinions about controversial topics. By capturing data from Twitter corresponding to the related hashtags, #Abortion and #Brexit, we primarily analyzed tweets discussed about the topics and created time series graph to identify polarization of people's opinions. However, we observed that neither a long-term discussed topic like #Abortion nor a topic whose result has been obtained like #Brexit, has significant changes in sentiments recently. By considering only in 2016, we found that both topics had similar behavior. The number of tweets remained steady from the beginning of the year, then rose markedly in August, the similar pattern can be seen in Fig.5. #Abortion #Brexit Month Positive Negative Positive Negative Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

7 Table 1. Number of positive tweets and negative tweets in 2016 separated by topic Fig. 5. No of Tweets in 2016 separated by topic In order to explain such coincidence we looked into individual tweets. Perhaps this is because the result of the Brexit referendum has been addressing abortion issues leaving the European Union is viewed as crisis of immigration for Irish woman. This seemed to cause more negative reactions for against in each topic. By calculating the ratio of Negative tweet to Positive tweets, it shows incremental trends in negative opinions. See in Fig.6 Fig. 6. Increasing of Negative tweet in 2016 In this work, we have studied one aspect of polarization. Our work was based on an assumption that people who give many opinions in two-side debate topic are likely to be aggressive. By means of sentiment analysis, we performed the primary experiment that may aid in making sense of sentiment changes in Twitter message discussed about controversial topics. We extracted the sentiment polarity for a long time debating issue in US, i.e. Abortion, and a national political decision in UK, i.e. Brexit. The experimental results of the average positive and negative sentiment polarity towards such topics did not show the clear support on the assumption, only the slight polarization was observed in one of our two topics studied. In short, the twitter users did not show any radicalization as a whole at least for the two topics studied. However, we may need to look into the data to examine the polari- 173

8 ty change of individuals as well as checking whether a small part of users have been radicalized or not. A limitation of this approach is that we need at least several hundreds of tweets a year. However, the tweets do not represent everyone who watched the debate, only those who had adopted Twitter and had chosen to respond. Measuring population sentiment from system like Twitter could not be substituted for a real poll. Having knowledge about user s background would better to see sentiment response of population biases [11] Based on these primary results, future work will focus on improving lexicon for a given topics. Also determining sentiment strength on the polarity classification. References 1. Pang, B., Lee, L., & Vaithyanathan, S. (2002, July). Thumbs up?. sentiment classification using machine learning techniques. In Proceedings of the ACL-02 conference on Empirical methods in natural language processing-volume 10 (pp ). Association for Computational Linguistics. 2. Conover, M. D., Gonçalves, B., Flammini, A., & Menczer, F. (2012). Partisan asymmetries in online political activity. EPJ Data Science, 1(1), Adamic, L. A., & Glance, N. (2005, August). The political blogosphere and the 2004 US election: divided they blog. In Proceedings of the 3rd international workshop on Link discovery (pp ). ACM. 4. Barnaghi, P., Ghaffari, P., & Breslin, J. G. (2016, March). Opinion Mining and Sentiment Polarity on Twitter and Correlation Between Events and Sentiment. In Big Data Computing Service and Applications (BigDataService), 2016 IEEE Second International Conference on (pp ). IEEE. 5. Smith, L. M., Zhu, L., Lerman, K., & Kozareva, Z. (2013, September). The role of social media in the discussion of controversial topics. In Social Computing (SocialCom), 2013 International Conference on (pp ). IEEE. 6. Gao, H., Mahmud, J., Chen, J., Nichols, J., & Zhou, M. X. (2014, May). Modeling User Attitude toward Controversial Topics in Online Social Media. In ICWSM. 7. Hasan, K. S., & Ng, V. (2013). Stance Classification of Ideological Debates: Data, Models, Features, and Constraints. In IJCNLP (pp ). 8. Thomas, M., Pang, B., & Lee, L. (2006, July). Get out the vote: Determining support or opposition from Congressional floor-debate transcripts. In Proceedings of the 2006 conference on empirical methods in natural language processing (pp ). Association for Computational Linguistics. 9. Miller, G. A., Beckwith, R., Fellbaum, C., Gross, D., & Miller, K. J. (1990). Introduction to WordNet: An on-line lexical database. International journal of lexicography, 3(4), Andrea Esuli, Fabrizio Sebastiani, "SENTIWORDNET: A Publicly Available Lexical Resource for Opinion Mining". In Proceedings of the 5th Conference on Language Resources and Evaluation (LREC 06), Pg Diakopoulos, Nicholas A., and David A. Shamma. "Characterizing debate performance via aggregated twitter sentiment." Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. ACM,