Political Talk Show Videos as Data

Videos have become the most consumed medium of political communication. Though, computational methods for video analysis remain relatively underexplored despite their significant potential. We contribute to fill this methodological gap by discussing the use of machine learning tools to extract data from videos. We present a case study where we construct a granular second-by-second dataset of over 2,000 Political Talk Show episodes. We show how to extract information regarding who is speaking, when they are speaking, what they are speaking about, and how they are speaking (tone of the voice and face emotion). Using these data, we provide a descriptive analysis of the trends observed in terms of representation, unveiling overt but also subtle forms of slant. We then use minute-by-minute viewership data to describe how different settings are related to changes in viewership, showing that heated discussion, and the presence of non-politicians’ speakers tend to garner higher audience.