Text as Actuator: Text-Driven Response Modeling and Prediction in Politics

<p dir="ltr">In this work, we develop a series of prediction tasks on ”actuating text”, defined as text that evokes– or is written to evoke– responses from its readership. Examples include blog posts with reader comments and product reviews with social tagging or ratings. Some traditional text collections, such as legislative proceedings, can also be seen as varieties of this type of text: legislative actions associated with a bill are, after all, the legislature’s collective reaction to the original text. </p><p dir="ltr">This thesis examines response prediction tasks in two distinct domains in contemporary U.S. politics: the political blogosphere and the United States Congress. In the blogosphere, we examine the relation between political topics and user comments they generate among the highly partisan readership community. In the U.S. Congress, we examine how bills survive the congressional committee system, a highly selective scrutinizing phase that happens before the general voting, and the relationship between a congress person’s microblog messages and the campaign contribution they receive from interest groups. </p><p dir="ltr">We propose several probabilistic models to predict attributes of the responses based on statis tical analysis of the associated texts. We anticipate that such models will ultimately prove useful in user assistive applications like recommendation and filtering systems, or serve as a technique to gather critical intelligence for scholars, content providers, or whoever is interested in learning the response trends among the target populations. In addition to achieving high prediction accuracy, we aim to shed light on the underlying response process, thereby contributing to social and political science research.</p>