A Frame of Mind: Using Statistical Models for Detection of Framing and Agenda Setting Campaigns

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 1629–1638, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics A Frame of Mind: Using Statistical Models for Detection of Framing and Agenda Setting Campaigns Oren Tsur Harvard University & Northeastern University orentsur@seas.harvard.edu Dan Calacci Northeastern University dcalacci@ccs.neu.edu David Lazer Northeastern University & Harvard University d.lazer@neu.edu Abstract Framing is a sophisticated form of dis- course in which the speaker tries to in- duce a cognitive bias through consis- tent linkage between a topic and a spe- cific context (frame). We build on po- litical science and communication theory and use probabilistic topic models com- bined with time series regression analy- sis (autoregressive distributed-lag models) to gain insights about the language dy- namics in the political processes. Pro- cessing four years of public statements is- sued by members of the U.S. Congress, our results provide a glimpse into the com- plex dynamic processes of framing, atten- tion shifts and agenda setting, commonly known as ‘spin’. We further provide new evidence for the divergence in party disci- pline in U.S. politics. 1 Introduction Language is one of the main tools used by politi- cians to promote their agenda, gain popularity, win elections and drive societal change (Luntz, 2007). The growing availability of online archives of po- litical data such as public statements, bill pro- posals, floor speeches, interviews or social me- dia streams allows computational analysis of many aspects of the political process. The analysis performed can increase transparency, facilitate a better educated constituency and improve under- standing of the political process. In this paper we propose a framework for au- tomatic analysis of a large collection of political texts. Specifically, we demonstrate how the use of Bayesian methods and time series analysis cap- tures the different ways in which political parties control the political discourse. We show that topic ownership and framing strategies can be inferred using topic models. Moreover, we demonstrate how the models learned are used to construct time series of expressed agendas. These time series are fitted using autoregressive distributive-lag models in order to learn the partisan temporal relations be- tween topics and expressed agendas. This framework could also be applied in other domains such as ideology divergence in online forums of radical groups or for measuring the changes in public sentiment toward commercial brands. Contribution (i) To the best of our knowledge this is the first work to analyze framing strategies on large scale in an unsupervised manner1. (ii) we combine topic models with regression analy- sis in recovering longitudinal trends. (iii) We fur- ther provide evidence for the dynamics of framing campaigns, commonly known as ‘political spin2’. Finally, (iv) we show how this framework can shed new light on the broad scholarship on the diver- gence of party discipline. 2 Related Work 2.1 Political Communication Theory Some of the theoretical constructs employed by Political Science scholars to describe features of the political …

Content truncated to 500 words