Persuasion, Ideology, and Speech: Using automated text analysis to model opinion formation and change

Political opinion forms in the torrent of language surrounding all of us, but despite the recent proliferation of speech and social network data, existing models of political persuasion and communication remain rooted in simplistic frameworks of low-dimensional ideological spaces and basic expressive or information-transmitting speech. This work instead develops new techniques in automated text analysis and large-scale causal inference to model the complex interplay between speech, belief, and political behaviors. It begins with the simple case of expressive speech in legislatures, where new Bayesian and spatial text scaling methods allow us to accurately predict both the ideology and voting behavior of individual legislators. In the more complex case of persuasive speech, ensembles of simple regression models allow us to measure the effects of hundreds of different TV ads, and new text methods allow us to predict the effects of ads based only on their text, and to infer a dense interplay of persuasive strategies not captured by existing approaches. The work finishes with the most complex and common case of persuasive speech, interpersonal conversation. Employing a new dataset of millions of online political discussions, a new semi-supervised application of a Bayesian topic model is used to reveal the conceptual network of ideas and issues at work in the minds of discussants; to predict the topics of conversational replies based on what was said before; and to infer the long-term effects of this speech via a panel vector autoregression model. This approach to modeling persuasive speech should also be broadly applicable to many other political domains: the media, the judiciary, Congressional bills and committees, and the burgeoning world of social networks.