Redalyc.A Review on Political Analysis and Social Media

Procesamiento del Lenguaje Natural ISSN: 1135-5948 [email protected] Sociedad Española para el Procesamiento del Lenguaje Natural España

Vilares, David; Alonso, Miguel A. A review on political analysis and social media Procesamiento del Lenguaje Natural, núm. 56, 2016, pp. 13-23 Sociedad Española para el Procesamiento del Lenguaje Natural Jaén, España

Available in: http://www.redalyc.org/articulo.oa?id=515754423001

How to cite Complete issue Scientific Information System More information about this article Network of Scientific Journals from Latin America, the Caribbean, Spain and Portugal Journal's homepage in redalyc.org Non-profit academic project, developed under the open access initiative Procesamiento de Lenguaje Natural, Revista nº 56, marzo de 2016, pp 13-23 recibido 28-10-2015 revisado 15-01-2016 aceptado 01-02-2016

A review on political analysis and social media ∗

Una revisio´n del an´alisis pol´ıtico mediante la web social

David Vilares y Miguel A. Alonso Grupo LyS, Departamento de Computacioń, Universidade da Corunã Campus de A Corunã, 15071 A Corunã, Spain {david.vilares, miguel.alonso }@udc.es

Resumen: En los pa´ıses democráticos, conocer la intenciónde voto de los ciu- dadanos y las valoraciones de los principales partidos y l´ıderes pol´ıticos es de gran interéstanto para los propios partidos como para los medios de comunicacióny el pu´blico en general. Para ello se han utilizado tradicionalmente costosas encuestas personales. El auge de las redes sociales, principalmente Twitter, permite pensar en ellas como una alternativa barata a las encuestas. En este trabajo, revisamos la bibliograf´ıa cient´ıfica más relevante en este a´mbito, poniendo especial énfasis en el caso espanõl. Palabras clave: Análisis pol´ıtico, Análisis del sentimiento, Twitter Abstract: In democratic countries, forecasting the voting intentions of citizens and knowing their opinions on major political parties and leaders is of great interest to the parties themselves, to the media, and to the general public. Traditionally, expensive polls based on personal interviews have been used for this purpose. The rise of social networks, particularly Twitter, allows us to consider them as a cheap alternative. In this paper, we review the relevant scientific bibliographic references in this area, with special emphasis on the Spanish case. Keywords: Political Analysis, Sentiment Analysis, Twitter

1 Introduction of political analysis on Twitter, most research has focused on predicting electoral outcomes, The adoption of social media and its use for although Twitter is also a valuable tool for widespread dissemination of political infor- tasks such as identifying the political pref- mation and sentiment is so remarkable that erences of the followers of an account (Gol- it has impacted traditional media. Nowa- beck and Hansen, 2011) and monitoring day- days, Twitter is a convenient tool for journal- to-day change and continuity in the state of ists in search of quotes from prominent news an electoral campaign (Jensen and Anstead, sources, e.g., politicians (Lassen and Brown, 2013; Wang et al., 2012). 2011), as they can add direct quotes to sto- ries without having the source in front of a In this article, we review the relevant sci- microphone or camera (Broersma and Gra- entific literature dealing with Twitter as a ham, 2012). source for political analysis, with special em- Current computational techniques (Mo- phasis on the Spanish case. In Section 2 we hammad et al., 2015) make possible to auto- consider work focused on predicting electoral matically determine the sentiment (positive outcomes, while in Section 3 we consider that or negative) and the emotion (joy, sadness, work dealing with the political preferences of etc.) expressed in a tweet, the purpose be- individual users. In Section 4 we consider hind it (to point out a mistake, to support, the use of Twitter as a forecasting tool in the to ridicule, etc.) and the style of writing Spanish political arena. Conclusions are pre- (statement, sarcasm, hyperboles, etc.). As sented in Section 5. a result, a lot of research activity has been devoted to analyze social media. In the field 2 Predicting electoral outcomes One of the first studies on the prediction ∗ This research is partially supported by Ministerio de Econom´ıay Competitividad (FFI2014-51978-C2). of electoral outcomes was performed by Tu- David Vilares is partially funded by the Ministerio de masjan et al., (2010). They analyze 104 003 Educacioń, Cultura y Deporte (FPU13/01180) Twitter messages mentioning the name of at ISSN 1135-5948 © 2016 Sociedad Española para el Procesamiento del Lenguaje Natural David Vilares, Miguel A. Alonso least one of the six parties represented in the ately large presence in social media, and a German parliament or prominent politicians lower result for Fianna Fáil, a party that at- of these parties, that were published in the tracted a low volume of tweets and plenty weeks leading up to the 2009 federal election of negativity, however it is traditionally the of the German parliament. The messages largest Irish party and thus it enjoys a degree were downloaded in German and automati- of brand loyalty. cally translated into English to be processed In a similar line, Effing, van Hillegersberg, by LIWC (Pennebaker, Francis, and Booth, and Huibers (2011) test whether there ex- 2001). They show that the mere number of ists a correlation between the use that Dutch tweets mentioning parties or politicians re- politicians made of social media and the in- flect voters preferences and comes close to dividual votes. Their study concludes that traditional election polls. They find surpris- the results of national elections where corre- ing that, although only 4% of users wrote lated with the compromise of politicians with more than 40% of messages, these heavy social media, but the same was not true for users were unable to impose their political local elections. One of the novelties of the opinion on the discussion, a fact they at- study is the introduction of an standarized tribute to the large number of participants on framework, Social Media Indicator (SMI), for Twitter who make the information stream as measuring the participation of politicians and a whole more representative of the electorate. how they interact with the public. Therefore, the main conclusion of Tumasjan O’Connor et al., (2010) try to determine if et al., (2010) is that Twitter may comple- the opinions extracted from Twitter messages ment traditional polls as political forecasting about the US presidential approval and the tool, although they also point out several lim- 2008 US presidential elections, correlate the itations of their approach: a Twitter sample opinions obtained by means of classical polls. may not be representative of the electorate, They collect messages over the years 2008 replies to messages in the sample that do not and 2009 and derive day-to-day sentiment mention any party or politician may be rele- scores by counting positive and negative mes- vant but they are missed, the dictionary may sages: a message is defined as positive if it be not well-tailored for the task, and the re- contains any positive word, and negative if sults may be not generalizable to other spe- it contains a negative one (a message can be cific political issues. both positive and negative). With this simple Bermingham and Smeaton (2011) use the sentiment analysis technique, they find many 2011 Irish General Election as a case of study, examples of falsely detected sentiment, but collecting 32 578 tweets relevant to the five they consider that, with a fairly large num- main Irish parties, where relevance is defined ber of measurements, these errors will cancel by the presence of the party names and their out relative to the aggregate public opinion. abbreviations, along with the election hash- They also find that recall is very low due to tag #ge11, with tweets reporting poll results the lexicon, designed for standard English. being discarded. They apply a volume-based To make predictions, day-to-day sentiment is measure defined as the proportional share volatile, so smoothing is a critical issue in or- of mentions for each party, and sentiment der to force a consistent behavior to appear analysis measures that represent the share over longer periods of time. Finally, they find of tweets with positive and negative opinion the sentiment rate correlated the presidential and, for each party, the log-ratio of sentiment approval polls, but it does not correlate to for the tweets mentioning it. They find that the elections polls. Unlike Tumasjan et al., the best method for predicting election re- (2010), they find that message volume has sults is the share of volume of tweets that a not a straightforward relationship to public given party receives in total, followed closely opinion. For the same 2008 US presidential by the share of positive volume. However, elections, Gayo-Avello (2011) collects 250 000 the mean absolute error of 5.85% is signif- Twitter messages published by 20 000 users in icantly higher than that achieved by tradi- seven states, finding that the correlation be- tional polls. Examining the errors, they find tween population and number of tweets and that they forecasted a higher result for the users was almost perfect. He applies four Green party, whose supporters tend to be simple sentiment analysis techniques to that more tech-savvy and have a disproportion- collection that also fail to predict the elec- 14 A review on political analysis and social media tion outcomes, concluding that the prediction negative opinions than Rommey. On the con- error is due to younger people is overrepre- trary, in the case of French Presidential elec- sented in Twitter, and that Republican sup- tion, the elected President Holland has less porters had tweeted much less than Demo- tweets than Sarkozy (incumbent candidate) cratic voters. with positive and neutral opinions but also DiGrazia et al., (2013) analyze 542 969 much less tweets with negative opinions. tweets mentioning candidates as well as data Caldarelli et al., (2014) monitor 3 million on elections outcomes from 795 competitive tweets during the 2013 General election in races in the 2010 and 2012 US Congressional Italy in order to measure the volume of tweets Elections and socio-demographic and control supporting each party. In this election, the variables such as incumbency, district par- three major parties got a similar number of tisanship, median age, percent white, per- votes but few traditional polls were able to cent college educated, median household in- predict the final outcomes. Although the come, percent female and media coverage. tweet volume and time evolution do not pre- They show that there is a statistically signifi- cisely predicted the election outcome, they cant association between tweets that mention provided a good proxy of the final results, a candidate for the US House of Represen- detecting a strong presence in Twitter of the tatives and the subsequent electoral perfor- (unexpected) winner party and the (also un- mance. They also find that districts where expected) relative weakness of the party fi- their models under-perform tend to be rel- nally occupying the fourth position. They atively noncompetitive and that a few dis- find that predicting results for small parties tricts have idiosyncratic features difficult to is difficult, receiving a too large volume of model, such as a rural district that had voted tweets when compared to their electoral re- for a Democratic congressman while voting sults. Moreover, a relevant 7.6% of votes strongly for the Republican presidential can- went to very small parties which were not didate. They conclude that (1) social me- considered in their study. dia are a better indicator of political behav- Lampos, Preo¸tiuc-Pietro, and Cohn ior than traditional TV media; (2) they can (2013) propose an approach for filtering serve as an important supplement to tradi- irrelevant tweets from the stream in order to tional voter surveys; and (3) they are less accurately model the polls in their prediction likely to be affected by social desirability bias in voting intentions for the three major than polling data, i.e., a person who partici- parties in the United Kingdom and for the pates in a poll may not express opinions per- four major parties in Austria. Gaurav et al., ceived to be embarrassing or offensive but so- (2013) predict with a low error margin the cially undesirable sentiments are captured in winners of the Venezuelan, Parguayan and social media. Ecuatorian Presidential elections of 2013. Contractor and Faruquie (2013) try to The best results are attained with a volume- use Twitter to predict the daily approval based approach consisting of measuring the rating of the two candidates for the 2012 number of tweets mentioning the full names US presidential elections. They formulate of candidates or mentioning the aliases of the issue as a time series regression problem candidates jointly with a electoral keywork. where the approval rate for each candidate is dependent on the bigrams (two consecu- 2.1 Controversy tive words) mentioned in messages written As a consequence of the mixed results ob- by his supporters. They find that 227 bi- tained in these studies, some authors are grams were causal for the Democratic candi- skeptics about the feasibility of using Twitter date and 183 bigrams for the Republican can- to predict the outcomes of electoral processes. didate. Nooralahzadeh, Arunachalam, and Jungherr, Ju¨rgens, and Schoen (2012) ar- Chiru (2013) compare the sentiment that pre- gue that, taking into account all of the par- vailed before and after the presidential elec- ties running for the elections, and not only tions taking place in 2012 in USA and France. the six ones with seats in the German parlia- In the case of the US Presidential election, ment, the approach of Tumasjan et al., (2010) they find that there are more tweets relat- would actually have predicted a victory of the ing Obama (incumbent candidate) with posi- Pirate Party, which received a 2% of the votes tive and neutral opinions and less tweets with but no seats in the parliament. 15 David Vilares, Miguel A. Alonso

Gayo-Avello (2012) indicates that senti- be predicted from the interaction with po- ment analysis methods based on simplis- litical parties. For this purpose, they build tic assumptions should be avoided, devot- an interaction profile for each party as a lan- ing more resources to the study of senti- guage model from the content of the tweets ment analysis in politics before trying to pre- by the party candidates, and the preference dict elections. Moreover, Metaxas, Musta- of a user is assessed according to the align- faraj, and Gayo-Avello (2011) find that elec- ment of user tweets with the language mod- toral predictions on Twitter data using the els of the parties. Their method is evalu- published research methods at that time are ated on a set of users whose political pref- not better than chance and that even when erences are known based on explicit state- the predictions are better than chance, as ments made on election day or soon after, when they were applied to a corpus of mes- in the context of Alberta 2012 general elec- sages during the 2010 US Congressional election. They find that, although less precise tions (Gayo-Avello, Metaxas, and Musta- than humans, for some parties their method faraj, 2011), they were not competent com- outperforms human annotators in recall, and pared to the trivial method of predicting revealed that politically active users are less through incumbency given that current hold- prone to change their preferences than the ers of a political office tends to maintain rest of users. Pennacchiotti and Popescu the position in an electoral process. To (2011) try to classify 10 338 Twitter users as corroborate their statement, they apply a being either Democrats or Republicans, find- lexicon-based sentiment analysis technique ing that the linguistic content of the user’s to a dataset of Twitter data compiled dur- tweets is highy valuable for this task, while ing the 2010 US Senate special election in social graph information has a negligible im- Massachusetts (Chung and Mustafaraj, 2011) pact on the overall performance. and they find that, when compared against Monti et al., (2013) analyze the phe- manually labeled tweets, its accuracy is only nomenon of political disaffection in Italy, i.e., slightly better than a classifier randomly as- negative sentiment towards the political sys- signing the three labels of positive, negative tem in general, rather than towards a par- and neutral to Twitter messages. ticular politician, policy or issue. For this On the other hand, Huberty (2013) points purpose, they apply sentiment analysis tech- out that US elections pose a very high bar, niques on political tweets to extract those since forecasts must beat the simple heuris- with negative sentiment, to then select the tic of incumbency that reliably predicts fu- tweets that refer to politics in general rather ture winners with high accuracy, even in os- than specific political events of personalties. tensible competitive races. He also finds They find a strong correlation between their that algorithms trained on one election for results and political disaffection as measured the U.S. House of Representatives perform in public opinion surveys. They also show poorly on a subsequent election, despite hav- that important political news of Italian news- ing performed well in out-of-sample tests on papers are often correlated with the highest the original election. peaks of disaffection. Prasetyo and Hauff (2015) point out that There are great difference in how electoral traditional polls in developing countries are processes are driven in developed and devel- less likely to be reliable than in developed oping countries. In this respect, Razzaq, Qa- countries, therefore they often result in a high mar, and Bilal (2014) use sentiment analy- forecasting error. Taking the 2014 Indone- sis to study the Twitter messages related to sian Presidential Election as a case study, the 2013 Pakistan general election, finding they show that a Twitter prediction based on there are two groups of users, one formed by sentiment analysis outperformed all available people living outside Pakistan and that only traditional polls on national level. could participate in political discussion in social media, and a second group of users living 3 Predicting political preferences in Pakistan. In this latter group, they also The aim of the work of Makazhanov and observed differences, both in volume and sen- Rafiei (2013) is not to forecast election out- timent, among users living in large cities and comes, but to predict the vote of individual in rural areas with low literacy rates. Fink et users, arguing that political preference can al., (2013) analyze the 2011 Nigerian Presi- 16 A review on political analysis and social media dential election and find that volume counts published by 380 164 distinct users during the of the mentions of the two major candidates same election, concluding again that mem- correlates strongly with polling and election bers of political parties tend to almost ex- outcomes, but that other candidates are over- clusively propagate content created by other represented. However, the particular ethnic members of their own party. They also ob- divide of Nigerian population makes religion serve that politicians conceive Twitter more the best predictor of electoral support, with as a place to diffuse their political messages place of living and education as significant than to enage in conversations with citizens, predictors as well. although minor and new parties are more prone to exploit the communication mecha- 4 Twitter as a tool for political nisms offered by Twitter; and that messages analysis in Spain corresponding to the winner party become With respect to the analysis of messages more and more positive until election day. regarding the political situation in Spain, Barberáand Rivero (2012) find that in Penã-López, Congosto, and Aragón (2011) the political debate in Twitter in Spain, 65% study networked citi zen politics, in particu- of participants are men compared to 35% lar the relations among the Spanish indig- of women and that the geographical distri- nados movement, traditional political par- bution of users corresponds to the distri- ties and mass media. Criado, Mart´ınez- bution of population in the country, except Fuentes, and Silván(2013) note the high de- that Madrid is overrepresented, with no sig- gree of use of Twitter as a channel of po- nificant differences between the behavior of litical communication during electoral peri- those living in large cities and in the rest of ods. Congosto, Fernańdez, and Moro Egido Spain. They also find a strong polarization (2011) corroborate results obtained for other of the political debate, since those citizens countries (Livne et al., 2011; Conover et al., with a stronger party identification monopo- 2012; Conover et al., 2011) that observed how lize much of the conversation, with the com- Twitter users are grouped by political affin- munication related to PP being highly struc- ity when transmitting information. A simi- tured and hierarchical, while the communi- lar grouping by ideological reasons is found cation concerning the PSOE is much more by Romero-Fr´ıas and Vaughan (2012) among horizontal and interactive. Spanish political parties and traditional mass In this context, assuming that individu- media when analyzing the linking practices als prefer to follow Twitter users whose ide- in the websites of both kinds of organiza- ology is close to theirs, Barberá(2012) con- tion, with left-wing media closer to PSOE siders the ideology or party identification of (socialist party) and right-wing media closer a Twitter user as a latent variable that can- to PP (conservative party). In the same not be observed directly, proposing the use of line, Romero-Fr´ıas and Vaughan (2010) find Bayesian inference to derive it from the list that ideology was the main factor in the clus- of accounts that each user follows. He takes tering of European political parties belonging as seeds the accounts of the top 50 politi- to the, at that time, 27 member states of the cians from PP and PSOE with the highest European Union, followed by country or re- number of followers. Then, he applies his ap- gional affiliation. proach on a random sample of 12 000 active Borondo et al., (2012) also find that politi- users during the 2011 Spanish elections. He cians mentioned and retweeted mostly their tries to validate the technique by consider- own partisans, after analyzing 370 000 Twit- ing as additional seeds the official accounts of ter messages written by over 100 000 users other two minority parties (IU and UPyD), during the 2011 Spanish general Election, obtaining inconclusive results that seems to where half of the messages were posted by support the idea that the latent variable is only 7% of participants, just 1% of users were not measuring ideology but rather a combi- the target for half of the mentions, 78% of nation of both policy preferences and party mentions were for politicians, 2% of the users support. In order to determine whether the causes half of the retweets and the source approach places both politicians and citizens of 63% of the retweeted messages were cre- on the same scale, he applies a lexicon-based ated by mass media accounts. Aragoń et sentiment analysis technique, observing that al. (2013) analyze 3 074 312 Twitter messages socialist candidates attain a better average 17 David Vilares, Miguel A. Alonso evaluation among left-wing Twitter users and tions in 2012. They compute the support that conservative candidates attain a bet- of each party by the number of followers of ter evaluation among right-wing users, as ex- the Twitter accounts of political parties and pected. A similar correlation is found be- their leaders. For the two major parties, PP tween the value of the latent variable for each and PSOE, the results computed by Deltell, user and the support of hashtags promoted Claes, and Osteso (2013) are closer to the fi- by the socialist and conservative parties. nal election outcomes than traditional polls. Borondo et al., (2012) confirm the finding So, for the PP they predict a 40.48% of votes of Tumasjan et al., (2010) that there exists for a 40.66% final result, while for PSOE they a correlation between the number of times a predict a 36.31% of votes for a final score political party is mentioned on Twitter and of 39.52%. We must point out that tradi- the electoral outcomes, but they only con- tional polls failed in most predictions: al- sider parties that obtained more than 1% of though they predicted rightly that PP would votes. Deltell (2012) analyzes the presence have more votes than PSOE, they predicted of one of these minor parties, eQuo, on so- a 10% difference between them when in the cial media during the 2011 Spanish General end it was less than 2%. This situation al- Election to question the efficiency and effec- lowed the leader of PSOE to be elected as tiveness of social networks in the modifica- regional president with the support of the tion of the vote and in predicting election re- elected parliamentarians of IU. The authors sults. At that time, eQuo was a newly cre- confirm that their method is not accurate ated green party, without enough budget for for small or newly created political parties, a conventional electoral campaign, thus, no in particular, they were completely wrong in TV, radio or newspapers ads were possible. predicting the votes for IU, which they at- In addition, as a completely new party, no tributed to the low activity on Twitter of IU’s free airtime was granted on public TV and leader. radios, and any privately-owned media of- Cotelo et al., (2015) use the follower- fered significant coverage. As a result, the folowee relationship to cluster politically ac- electoral strategy of eQuo was based mainly tive users in Twitter. This information is on social media: for several days its propos- combined with the textual content of tweets als were trending topics on Twitter, its Face- in order to determine the sentiment (posi- book page was the most-visited and had more tive, negative or nautral) expressed in a given “likes” than the page of any other political tweet with respect to PP and PSOE, attain- party, and it was the party with more pres- ing a 75% accuracy. ence on YouTube. However, this apparently Vilares, Thelwall, and Alonso (2015) an- successful campaign on social media was not alyze the sentiment of 2 704 523 tweets refer- reflected in its electoral outcome, as the num- ring to Spanish politicians and parties from 3 ber of votes was so small than no represen- December 2014 to 12 January 2015. They de- tative was assigned to eQuo in the parlia- sc ribe the Spanish version of SentiStrength 1, ment. Surprisingly, the best results for eQuo an algorithm designed originally for analyz- were obtained in those districts in which this ing the sentiment of English texts in so- party was present physically by means of cial media (Thelwall, Buckley, and Paltoglou, traditional activities such as meetings, past- 2012), and how their sentiment scores are ing campaign posters, and recreational activ- used to build ranks for the politicians and ities. The interesting point here is that, dis- their parties, giving popularity ratings that regarding eQuo, simple methods relying on are comparable with those provided by the the number of Twitter followers and Face- classic polls, although tweet volume was a books “likes” seems to be reliable indicators much better predictor of voting intentions. of outcomes for the 2011 Spanish elections. A deeper analysis of politicians that had sen- Deltell, Claes, and Osteso (2013) study timent scores and that did not match those the political campaign on Twitter during the of their parties, suggested that these had at- Andalusian regional elections of 2012. They tracted negative media publicity that had focus on monitoring the Twitter profiles of been amplified in Twitter. Thus, Twitter re- the six most-voted political parties in An- sults may be useful to analyze the trajecto- dalusia in the Spanish elections of 2011 and their leaders for Andalusian regional elec- 1http://sentistrength.wlv.ac.uk/#Non-English 18 A review on political analysis and social media ries of individual politicians and to evaluate pects. Then, some clusters are grouped to- the impact of negative press coverage on their gether attending to the left vs. right political popular perception. dimension. The deep learning approach tried The task 4 of the TASS 2013 compe- by Vilares et al., (2015), based on LSTM tition (Villena-Román and Garc´ıa-Morera, recurrent neural networks, did not outper- 2013) consisted of classifying the political formed well-established machine learning ap- tendency of public figures (not necessarily proches, probably due to unsupervised pre- politicians) into one of four wings: left, cen- training and sentiment-specific were not con- ter, right or neutral. A training set was not sidered. provided, so participant teams need to de- fine their own strategies to categorise the au- 5 Conclusion and future work thors. This was a controversial issue since Over the last five years a lot of studies have the same party might belong to a different been conducted on the use of Twitter as a wing depending on their statutes or the polls cheap replacement for expensive polls involv- made to citizenship. The best performing ing personal interviews. Some initial satisfac- system (Pla and Hurtado, 2013) considered tory results were followed by disappointing a number of entities related with the main ones, sparking controversy over the method- political parties, which were classified into ology used and the management of biases in- one of the four proposed classes. If the mes- troduced by the demographics of active users sages of a user containing one of those en- on Twitter. However, we can see how in re- tities tend to be negative the user is prone cent years Twitter has been accepted as a to be against that political orientation, and valid tool of political analysis, although there vice versa. The task 3 of this same workshop are still problems to be solved, such as the was related with politics too: given a tweet handling of very small parties with very ac- where a representation of an entity (one of tive Twitter users; the management of small the four main national parties in 2013) oc- constituencies; the detection of spam pro- curs, participants where intended to classify duced by robots or users engaged in propa- the polarity of that entity. In this case, the ganda; and the treatment of countries with best performing system (Gamallo, Garc´ıa, multilingual population. and Fernández Lanza, 2013) assumed that the polarity of the whole tweets corresponded References to the polarity of the entity. Aragón, P., K.E. Kappler, A. Kaltenbrun- For the TASS 2015 competition (Villena- ner, D. Laniado, and Y. Volkovich. 2013. Románet al., 2015), the STOMPOL training Communication dynamics in Twitter dur- and test corpora were developed, formed by ing political campaigns: The case of the 784 and 500 tweets, respectively, related to 2011 Spanish national election. Policy and the six major Spanish political parties gath- Internet , 5(2):183–206, June. ered from 23rd to 24th of April 2015. For Barbera´, P. 2012. A new measure of party each tweet, the polarity of the aspects in- identification in Twitter. Evidence from volved (economy, health, education, political Spain. In 2nd Annual General Conference parties, other political issues) were identified. of the European Political Science Associa- Only three teams participated at TASS 2015 tion(EPSA) , Berlin, June. EPSA. task 2 (aspect-based sentiment analysis) on STOMPOL. The best-performing system ap- Barbera´, P. and G. Rivero. 2012. ¿Un tweet, plied a regular sentiment analysis system on un voto? Desigualdad en la discusión the context of each aspect under consider- pol´ıtica en Twitter. In I Congreso Inter- ation, defining context as a fixed-size win- nacional en ComunicaciónPol´ıtica y Es- dow around the aspect instance. Best re- trategias de Campanã, Madrid, July. AL- sults were attained by the approach by Park ICE. (2015), which clusters tweets according to Bermingham, A. and A.F. Smeaton. 2011. party-aspect pairs, on the assumptions that On using Twitter to monitor political sen- people who share similar preference or sta- timent and predict election results. In S. tus show similarity in the expression of sen- Bandyopadhyay and M. Okumurra, edi- timent and that people evaluate a political tors, Proceedings of the Workshop on Sen- party in multiple ways regarding different as- timent Analysis where AI meets Psychol- 19 David Vilares, Miguel A. Alonso

ogy (SAAIP 2011) , pages 2–10, Chiang Criado, J.I., G. Mart´ınez-Fuentes, and A. Mai, Thailand, November. Asian Federa- Silván. 2013. Twitter en campanã: las tion of Natural Language Processing. elecciones municipales espanõlas de 2011. Borondo, J., A.J. Morales, J.C. Losada, and Revista de investigaciones Pol´ıticas y So- R.M. Benito. 2012. Characterizing and ciológicas , 12(1):93–113. modeling an electoral campaign in the Deltell, L. 2012. Estrategias de comuni- context of Twitter: 2011 Spanish pres- caciónpol´ıtica en las redes sociales du- idential election as a case study. AIP rante la campanã electoral del 2011 en Chaos , 22(2):023138, June. Espanã: el caso de eQuo. In II Jornadas Broersma, M. and T. Graham. 2012. Social de la Asociación Madrilenã de Sociolog´ıa, media as beat. tweets as a news source Madrid, March. during the 2010 British and Dutch elec- Deltell, L., F. Claes, and J.M. Osteso. 2013. tions. Journalism Practice , 6(3):403–419. Predicciónde tendencia pol´ıtica por Twit- Caldarelli, G., A. Chessa, F. Pammolli, G. ter: Elecciones andaluzas 2012. Am- Pompa, M. Puliga, M. Riccaboni, and G. bitos: Revista internacional de comuni- Riotta. 2014. A multi-level geographi- cación , 22:91–100. cal study of Italian political elections from DiGrazia, J., K. McKelvey, J. Bollen, and Twitter data. PLOS ONE , 9(5):e95809, F. Rojas. 2013. More tweets, more May. votes: Social media as a quantitative in- Chung, J. and E. Mustafaraj. 2011. Can dicator of political behavior. PLOS ONE , collective sentiment expressed on Twitter 8(11):e79449, November. predict political elections? In Proceedings Effing, R., J. van Hillegersberg, and T. of the Twenty-Fifth Conference on Artifi- Huibers. 2011. Social media and politi- cial Intelligence (AAAI-11) , pages 1170– cal participation: Are Facebook, Twitter 1171, San Francisco, CA, August. AAAI. and YouTube democratizing our political Congosto, M.L., M. Fernández, and E. systems? In E. Tambouris, A. Macintosh, Moro Egido. 2011. Twitter y pol´ıtica: and H. de Bruijn, editors, Electronic Par- Información, opinión y ¿predicción? ticipation , volume 6847 of Lecture Notes Cuadernos de Comunicación Evoca , in Computer Science . Springer, Berlin and 4:11–15. Heidelberg, pages 25–35. Conover, M.D, B. Gonçalves, A. Flammini, Fink, C., N. Bos, A. Perrone, E. Liu, and J. and F. Menczer. 2012. Partisan asymme- Kopecky. 2013. Twitter, public opinion, tries in online political activity. EPJ Data and the 2011 Nigerian presidential elec- Science , 1(1):Article: 6, December. tion. In Proceedings of SocialCom / PAS- Conover, M.D., J. Ratkiewicz, M. Fran- SAT / BigData / EconCom / BioMed- cisco, B. Gonçalves, A. Flammini, and F. Com 2013 (SocialCom 2013) , pages 311– Menczer. 2011. Political polarization on 320, Washington, D.C., USA, September. Twitter. In Proceedings of the Fifth In- IEEE Computer Society. ternational AAAI Conference on Weblogs Gamallo, P., M. Garc´ıa, and S. and Social Media (ICWSM 2011) , pages Fernández Lanza. 2013. TASS: A 89–96, Barcelona, Spain, July. AAAI. naive-Bayes strategy for sentiment analy- Contractor, D. and T.A. Faruquie. 2013. Un- sis on Spanish tweets. In A. D´ıaz Esteban, derstanding election candidate approval I. Alegr´ıa Loinaz, and J. Villena Román, ratings using social media data. In WWW editors, XXIX Congreso de la Sociedad 2013 Companion , pages 189–190, Rio de Espanõla de Procesamiento de Lenguaje Janeiro, May. ACM Press. Natural (SEPLN 2013). TASS 2013 - Workshop on Sentiment Analysis at Cotelo, J.M., F. Cruz, F.J. Ortega, and J.A. SEPLN 2013 , pages 126–132, Madrid, Troyano. 2015. Explorando Twitter medi- Spain, September. ante la integración de informacioń estructurada y no estructurada. Procesamiento Gaurav, M., A. Srivastava, A. Kumar, and del Lenguaje Natural , 55:75–82, Septem- S. Miller. 2013. Leveraging candidate ber. popularity on Twitter to predict election 20 A review on political analysis and social media

outcome. In Proceedings of the 7th Work- of the Association for Computational Lin- shop on Social Network Mining and Anal- guistics , pages 993–1003, Sofia, Bulgaria, ysis (SNA-KDD 2013) , page Article No. August. ACL. 7, Chicago, IL. ACM. Lassen, D.S. and A.R. Brown. 2011. Twitter: Gayo-Avello, D. 2011. Don’t turn social The electoral connection? Social Sience median into another ‘literay digest’ poll. Computer Review , 29(4):419–436, Novem- Communications of the ACM , 54(10):121– ber. 128, October. Livne, A., M.P. Simmons, E. Adar, and L.A. Adamic. 2011. The party is over here: Gayo-Avello, D. 2012. No, you cannot pre- Structure and content in the 2010 elec- dict elections with Twitter. IEEE In- tion. In Proceedings of the Fifth Interna- ternet Computing , 16(6):91–94, Novem- tional AAAI Conference on Weblogs and ber/December. Social Media (ICWSM 2011) , pages 201– Gayo-Avello, D., P.T. Metaxas, and E. 208, Barcelona, Spain, July. AAAI. Mustafaraj. 2011. Limits of electoral pre- Makazhanov, A. and D. Rafiei. 2013. dictions using Twitter. In Proceedings of Predicting political preference of Twit- the Fifth International AAAI Conference ter users. In Proceedings of the 2013 on Weblogs and Social Media (ICWSM IEEE/ACM International Conference on 2011) , pages 490–493, Barcelona, Spain, Advances in Social Networks Analysis and July. AAAI. Mining (ASONAM 2013) , pages 298–305, Niagara, ON, Canada, August. ACM. Golbeck, J. and D.L. Hansen. 2011. Com- puting political preference among Twit- Metaxas, P.T., E. Mustafaraj, and D. Gayo- ter followers. In Proceedings of the Avello. 2011. How (not) to predict elec- SIGCHI Conference on Human Factors tions. In Proceedings of 2011 IEEE In- in Computing Systems (CHI 2011) , pages ternational Conference on Privacy, Secu- 1105–1108, Vancouver, BC, Canada, May. rity, Risk and Trust and IEEE Interna- ACM. tional Conference on Social Computing — PASSAT/SocialCom 2011— , pages 165– Huberty, M. 2013. Multi-cycle forecasting 171, Boston, Massachusetts, USA, Octo- of congressional elections with social me- ber. IEEE Computer Society. dia. In Proceedings of the 2nd workshop Mohammad, S.M., X. Zhu, S. Kiritchenko, on Politics, elections and data (PLEAD and J. Martin. 2015. Sentiment, emotion, 2013) , pages 23–29, San Francisco, CA, purpose, and style in electoral tweets. In- October. ACM. formation Processing and Management , Jensen, M.J. and N. Anstead. 2013. Psepho- 51(4):480–499, July. logical investigations: Tweets, votes and Monti, C., A. Rozza, G. Zapella, M. Zig- unknown unknowns in the republican nani, A. Arvidsson, and E. Colleoni. 2013. nomination process. Policy & Internet , Modelling political disaffection from Twit- 5(2):161–182, June. ter data. In Proceedings of the Sec- ond International Workshop on Issues of Jungherr, A., P. Ju¨rgens, and H. Schoen. Sentiment Discovery and Opinion Min- 2012. Why the pirate party won the Ger- ing (WISDOM 2013) , page Article No. 3, man election of 2009 or the trouble with Chicago, IL, USA. ACM. predictions: A response to Tumasjan, A., Sprenger, T. O., Sander, P. G., & Welpe, Nooralahzadeh, F., V. Arunachalam, and C. I. M. “Predicting Elections With Twitter: Chiru. 2013. 2012 presidential elections What 140 Characters Reveal About Polit- on Twitter — an analysis of how the ical Sentiment”. Social Science Computer US and French election were reflected in Review, , 30(2):229–234. tweets. In I. Dumitrache, A.M. Florea, and F. Pop, editors, Proceedings of the Lampos, V., D. Preo¸tiuc-Pietro, and T. 19th International Conference on Control Cohn. 2013. A user-centric model of Systems and Computer Science (CSCS voting intention from social media. In 2013) , pages 240–246, Bucharest, Roma- Proceedings of the 51st Annual Meeting nia, May. IEEE Computer Society. 21 David Vilares, Miguel A. Alonso

O’Connor, B., R. Balasubramanyan, B. Prasetyo, N.D. and C. Hauff. 2015. Twitter- Routledge, and N.A. Smith. 2010. From based election prediction in the develop- tweets to polls: Linking text sentiment ing world. In Proceedings of the 26th ACM to public opinion time series. In W.W. Conference on Hypertext and Social Media Cohen and S. Gosling, editors, Proceed- (HT 2015) , Middle East Technical Univer- ings of the Fourth International AAAI sity Northern Cyprus Campus, Cyprus, Conference on Weblogs and Social Media September. ACM. (ICWSM 2010) , pages 122–129, Washing- ton, DC, May. AAAI. Razzaq, M.A., A.M. Qamar, and H.S.M. Bi- lal. 2014. Prediction and analysis of Park, S. 2015. Sentiment classification us- pakistan election 2013 based on senti- ing sociolinguistic clusters. In J- Villena- ment analysis. In X. Wu, M. Ester, and Román, J. Garc´ıa-Morera, M. A.´ Garc´ıa- G. Xu, editors, Proceedings of the 2014 Cumbreras, E. Mart´ınez-Cámara, M.T. IEEE/ACM International Conference on Mart´ın-Valdivia, and L.A. Urenã-López, Advances in Social Networks Analysis and editors, TASS 2015 Workshop on Senti- Mining (ASONAM 2014) , pages 700–703, ment Analysis at SEPLN , volume 1397 of Beijing, China, August. IEEE. CEUR Workshop Proceedings , pages 99– Romero-Fr´ıas, E. and L. Vaughan. 2010. 104, Alicante, Spain, September. CEUR- European political trends viewed through WS.org. patterns of web linking. Journal of the Penã-López, I., M. Congosto, and P. Aragón. American Society for Information Science 2011. Spanish indignados and the evolu- and Technology , 61(10):2109–2121, Octo- tion of the 15M: Towards networked para- ber. institutions. In J. Balcells, A. Cerrillo- Romero-Fr´ıas, E. and L. Vaughan. 2012. Ex- i-Mart´ınez, M. Peguera, I. Penã-López, ploring the relationships between media M.J. Pifarré,and M. Vilasau, editors, Big and political parties through web hyper- Data: Challenges and Opportunities. Pro- link analysis: The case of Spain. Jour- ceedings od the 9th International Confer- nal of the American Society for Informa- ence on Internet, Law & Politics , pages tion Science and Technology , 63(5):967– 359–386, Barcelona, June. UOC-Huygens 976, May. Editorial. Thelwall, M., K. Buckley, and G. Paltoglou. Pennacchiotti, M. and A.-M. Popescu. 2011. 2012. Sentiment strength detection for the Democrats, Republicans and Starbucks social web. Journal of the American Soci- afficionados: User classification in Twit- ety for Information Science and Technol- ter. In KDD’11. Proceedings of the 17th ogy , 63(1):163–173, January. ACM SIGKDD international conference on Knowledge discovery and data mining , Tumasjan, A., T.O. Sprenger, P.G. Sand- pages 430–436, San Diego, CA, USA, Au- ner, and I.M. Welpe. 2010. Predicting gust. ACM. elections in Twitter: What 140 characters reveal about political sentiment. In Pennebaker, J.W., M.E. Francis, and R.J. W.W. Cohen and S. Gosling, editors, Pro- Booth. 2001. Linguistic inquiry and word ceedings of the Fourth International AAAI count: LIWC 2001 . Lawrence Erlbaum Conference on Weblogs and Social Media Associates, Mahway. (ICWSM 2010) , pages 178–185, Washing- ton, DC, May. AAAI. Pla, F. and L.-F. Hurtado. 2013. ELiRF- UPV en TASS-2013: Ana´lisis de sen- Vilares, D., Y. Doval, M.A. Alonso, and C. timientos en Twitter. In A. D´ıaz Es- Gómez-Rodr´ıguez. 2015. LyS at TASS teban, I. Alegr´ıa Loinaz, and J. Vil- 2015: Deep learning experiments for sen- lena Romań, editors, XXIX Congreso de timent analysis on Spanish tweets. In J. la Sociedad Espanõla de Procesamiento de Villena-Román, J. Garc´ıa-Morera, M. A.´ Lenguaje Natural (SEPLN 2013). TASS Garc´ıa-Cumbreras, E. Mart´ınez-Cámara, 2013 - Workshop on Sentiment Analysis M.T. Mart´ın-Valdivia, and L.A. Urenã- at SEPLN 2013 , pages 220–227, Madrid, López, editors, TASS 2015 Workshop on Spain, September. Sentiment Analysis at SEPLN , volume 22 A review on political analysis and social media

1397 of CEUR Workshop Proceedings , pages 47–52, Alicante, Spain, September. CEUR-WS.org. Vilares, D., M. Thelwall, and M.A. Alonso. 2015. The megaphone of the people? Spanish SentiStrength for real-time analysis of political tweets. Journal of Informa- tion Science , 41(6):799–813, December. Villena-Román, J. and J. Garc´ıa-Morera. 2013. TASS 2013 — workshop on sentiment analysis at SEPLN 2013: An overview. In A. D´ıaz Esteban, I. Alegr´ıa Loinaz, and J. Villena Romań, editors, XXIX Congreso de la Sociedad Espanõla de Procesamiento de Lenguaje Natural (SEPLN 2013). TASS 2013 - Workshop on Sentiment Analysis at SE- PLN 2013 , pages 112–125, Madrid, Spain, September. Villena-Román, J., J. Garc´ıa-Morera, M. A.´ Garc´ıa-Cumbreras, E. Mart´ınez-Cámara, M.T. Mart´ın-Valdivia, and L.A. Urenã- López. 2015. Overview of TASS 2015. In J. Villena-Román, J. Garc´ıa-Morera, M. A.´ Garc´ıa-Cumbreras, E. Mart´ınez- Cámara, M. T. Mart´ın-Valdivia, and L. A. Urenã-López, editors, TASS 2015 Work- shop on Sentiment Analysis at SEPLN , volume 1397 of CEUR Workshop Pro- ceedings , pages 13–21, Alicante, Spain, September. CEUR-WS.org. Wang, H., D. Can, A. Kazemzadeh, F. Bar, and S. Narayanan. 2012. A system for real-time Twitter sentiment analysis of 2012 U.S. presidential election cycle. In Proceedings of the 50th Annual Meeting of the Association for Computational Lin- guistics , pages 115–120, Jeju, Republic of Korea, July. ACL.