Ideological Phrase Indicators for Classification of Political Discourse
Total Page：16
File Type：pdf, Size：1020Kb
Ideological Phrase Indicators for Classification of Political Discourse Framing on Twitter Kristen Johnson, I-Ta Lee, Dan Goldwasser Department of Computer Science Purdue University, West Lafayette, IN 47907 john1187, lee2226, dgoldwas @purdue.edu { } Abstract 2015; Card et al., 2015; Baumer et al., 2015) as a way to automatically analyze political discourse Politicians carefully word their statements in congressional speeches and political news arti- in order to influence how others view an cles. Contrary to these sources, Twitter requires issue, a political strategy called framing. politicians to compress their ideas and reactions Simultaneously, these frames may also re- into 140 character long tweets. As a result, politi- veal the beliefs or positions on an issue cians have to cleverly choose how to frame contro- of the politician. Simple language fea- versial issues, as well as react to events and each tures such as unigrams, bigrams, and tri- other (Mejova et al., 2013; Tumasjan et al., 2010). grams are important indicators for identi- Framing decisions can be used to build support fying the general frame of a text, for both for political stances and they often reflect ideolog- longer congressional speeches and shorter ical differences between politicians. For example, tweets of politicians. However, tweets in debates concerning the issue of abortion, the may contain multiple unigrams across dif- stance opposing abortion is framed as “pro-life”, ferent frames which limits the effective- which reflects a moral or religious-based ideol- ness of this approach. In this paper, we ogy. Correctly identifying how issues are framed present a joint model which uses both lin- can help reveal the ideological base of the speaker. guistic features of tweets and ideological However, in many cases framing abstracts this in- phrase indicators extracted from a state-of- formation and groups content reflecting differing the-art embedding-based model to predict ideologies together under the same frame. As a the general frame of political tweets. concrete example consider the following tweets: 1 Introduction 1. POTUS exec. order on guns is a gross over- reach of power that tramples on the rights of Social media platforms have played an increas- law abiding Americans and our Constitution ingly important role in U.S. presidential elections, beginning in 2008. Among these, microblogs such 2. With this ruling #SCOTUS has upheld a critical as Twitter have a special role, as they allow politi- freedom for women to make their own decisions cians to react quickly to events as they unfold and about their bodies to shape the discussion of current political issues In both tweets, the same frame (Legality, Con- according to their views. stitutionality, & Jurisdiction) is used to discuss Framing is an important tool used by politicians two different issues: guns and abortion, respec- to bias the discussion towards their stance. Fram- tively. Despite the use of a similar frame, the two ing contextualizes the discussion by emphasizing tweets reflect opposing ideologies. specific aspects of the issue, which creates an as- A straight-forward approach for identifying sociation between the issue and a specific frame of these differences would be to refine the issue- reference. Research on issue framing in political independent general frames into more specific cat- discourse is rooted in social science research (Ent- egories. However, this would limit their general- man, 1993; Chong and Druckman, 2007) and re- ization and considerably increase the difficulty of cently has attracted growing interest in the natu- analysis, both for human annotators and for au- ral language processing community (Tsur et al., tomated techniques. Instead, we suggest to aug- 90 Proceedings of the Second Workshop on Natural Language Processing and Computational Social Science, pages 90–99, Vancouver, Canada, August 3, 2017. c 2017 Association for Computational Linguistics ment the frame analysis with additional informa- cators. We also include a qualitative analysis in tion. Our modeling approach is based on the ob- Section6 of several examples in which ideological servation that politicians often use slogans in both phrase indicators can help differentiate between their tweets and speeches. These are key phrases tweets with similar frame predictions that reflect used to indirectly indicate the political figures’ different ideologies. core beliefs and ideological stances. Identifica- tion of these phrases automatically decomposes 2 Related Work the frames into more specific categories. Previous computational works which analyze po- Consider the two tweets in the example above. litical discourse focus on opinion mining and In the first tweet, several phrases indicate the stance prediction from forums and tweets (Srid- frame: “exec. order”, “overreach of power”, har et al., 2015; Hasan and Ng, 2014; Abu- “rights of law abiding Americans”, “our constitu- Jbara et al., 2013; Walker et al., 2012; Abbott tion”. In the second tweet, the relevant phrases et al., 2011; Somasundaran and Wiebe, 2010, are “this ruling” and “upheld a critical freedom”. 2009; Johnson and Goldwasser, 2016; Ebrahimi All of these phrases indicate that the same frame is et al., 2016). A variety of social media based being used in both tweets. However, analyzing the predictions have been studied including: predic- specific terminology in each case and the context tion of political affiliation and other demograph- in which it appears helps capture the ideological ics of Twitter users (Volkova et al., 2015, 2014; similarities and differences. For example, in the Yano et al., 2013; Conover et al., 2011), pro- context of gun-rights debates, phrases highlight- file (Li et al., 2014b) and life event extraction (Li ing “law and order” and references to the constitu- et al., 2014a), conversation modeling (Ritter et al., tion tend to reflect a conservative ideology, while 2010), methods for handling unique microblog phrases highlighting upholding of freedoms in the language (Eisenstein, 2013), and the modeling abortion debate tend to reflect a liberal ideology. of social interactions and group structure in pre- Given the rapidly changing nature of trending dictions (Sridhar et al., 2015; Abu-Jbara et al., issues and political discourse on Twitter, our key 2013; West et al., 2014; Huang et al., 2012). technical challenge is to relay these ideological Works which focus on inferring signed social net- dimensions to an automated model, such that it works (West et al., 2014) and collective classifica- will be able to easily adapt to new issues and tion using PSL (Bach et al., 2015) are similar to language. Our model consists of two compo- the modeling approach of Johnson et al.(2017b), nents combined together: frame identification and which we extend in this paper. ideological-indicators identification. For the first Several previous works have explored framing piece we use a structured probabilistic model to in public statements, congressional speeches, and capture general framing dimensions by combining news articles (Fulgoni et al., 2016; Tsur et al., content and political context analysis. For the sec- 2015; Card et al., 2015; Baumer et al., 2015). ond task, we employ a state-of-the-art textual sim- Framing is further related to works which ana- ilarity model which captures and generalizes over lyze biased language (Recasens et al., 2013; Choi lexical indicators of key phrases that identify the et al., 2012; Greene and Resnik, 2009) and sub- politicians’ ideology. More details of both com- jectivity (Wiebe et al., 2004). Important to the ponents are described in Section4. language analysis of our work, Tan et al.(2014) In this paper we take a first step towards con- have shown how wording choices can affect mes- necting these two dimensions of analysis: issue sage propagation on Twitter. The study of political framing and ideology identification. We lay the sentiment analysis (Pla and Hurtado, 2014; Bakli- foundation for more advanced research by identi- wal et al., 2013), ideology measurement and pre- fying this connection, analyzing tweets authored diction (Iyyer et al., 2014; Bamman and Smith, by U.S congressional representatives, and extract- 2015; Sim et al., 2013; Djemili et al., 2014), poli- ing ideological phrase indicators. We build and cies (Nguyen et al., 2015), voting patterns (Gerrish analyze a joint model which combines the two di- and Blei, 2012), and polls based on Twitter polit- mensions. Our experiments in Section5 quantita- ical sentiment (Bermingham and Smeaton, 2011; tively compare the differences in frame prediction O’Connor et al., 2010; Tumasjan et al., 2010) are performance when using ideological phrase indi- also related to the study of framing on Twitter. 91 FRAME NUMBER,FRAME, AND BRIEF DESCRIPTION using 17 possible frames. A brief description of 1. Economic: Economic effects of a policy 2. Capacity & Resources: Resources lack or availability each frame is shown in Table1. 3. Morality & Ethics: Religious doctrine, righteousness, sense of responsibility Frame Overlap: Johnson et al.(2017b,a) found 4. Fairness & Equality: Distribution of laws, punish- that for most tweets, one or two frames were used. ments, resources, etc. among groups Additionally, in many cases, tweets authored by 5. Legality, Constitutionality, & Jurisdiction: Court cases and restriction and expressions of rights Republican and Democratic politicians use similar 6. Crime & Punishment: Crimes and consequences frames, both when discussing similar and