Topic Signatures in Political Campaign Speeches

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2342–2347 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Topic Signatures in Political Campaign Speeches Cl´ement Gautrais1, Peggy Cellier2, Ren´e Quiniou3, and Alexandre Termier1 1University of Rennes 1, IRISA, France 2INSA Rennes, IRISA, France 3Inria Rennes, IRISA, France Abstract Highlighting the recurrence of topics us- age in candidates speeches is a key fea- ture to identify the main ideas of each candidate during a political campaign. In this paper, we present a method combin- ing standard topic modeling with signa- ture mining for analyzing topic recurrence in speeches of Clinton and Trump during the 2016 American presidential campaign. The results show that the method extracts automatically the main ideas of each can- didate and, in addition, provides informa- tion about the evolution of these topics during the campaign. 1 Introduction Political discourse analysis (Van Dijk, 1998) is a branch of discourse analysis that aims at explicit- ing from speeches or debates the salient features of political discourses. From that point of view, a presidential election provides interesting datasets to study. Indeed, it is a major political event in a country and gives rise to many political meet- ings where candidates discuss personally selected societal problems and detail their own solutions. In that context, the identification of the favourite topics of candidates as well as how they evolve throughout the campaign is a crucial task. In (Savoy, 2010), the author presents an analy- sis of the evolution of topics in political speeches by comparing the words that are overused or un- derused by Obama and McCain during the 2008 US presidential campaign. The dynamics of these particular words usage is analyzed over monthly periods to identify the underlying dynamics of the campaign topics. A limitation of this approach is that the period is fixed (monthly) whereas pre- dictable (vote, debates) or unpredictable (scan- dals) events usually give the rhythm to a political campaign. Other work used topic modeling meth- ods, such as Latent Dirichlet Allocation (LDA) (Blei et al., 2003), Latent Semantic Anlayis (LSA) (Landauer et al., 1998) or Non-negative Matrix Factorization (NMF) (Lee and Seung, 1999), to study political texts (Prabhakaran et al., 2014; Quinn et al., 2010). For instance in (Quinn et al., 2010), a topic model for legislative speech is de- fined. However, those works study topics one at a time whereas a set of co-referenced topics is more relevant since it constitutes the core of a candi- date’s political program. There exist other works about political text analysis (Calvet and V´eronis, 2008) but they focus on the use of predefined single words in speeches over time, whereas we aim at finding (without knowing in advance which topic are recurrent) the recurrent topics (multiple topics) usage over time. In this paper, we propose to identify in political speeches the favourite topics considered by each candidate as well as how and when they evolve throughout the campaign. In our opinion, this gives critical clues …

Content truncated to 500 words