Talk: Sílvia Majó on Media Diets

silvia majo
Vvisiting student Sílvia Majó gave a talk today on “Media Diets and Contested Political Events”. Her paper was selected for presentation at the ICA 2016 conference “Communicating with Power”, which this year had an acceptance rate of 46%. Congratulations to Sílvia!

Talk: Nick Beauchamp visits DiMeNet

Nick Beauchamp came to DiMeNet today to give a talk on text analysis for research on persuasion and deliberation. The title and abstract of his talk are below.

Modeling Deliberation and Persuasion using Text Analysis

This project seeks to model political deliberation, persuasion, and opinion in its natural state: the free-form exchange of words and arguments. It begins with a new corpus of political discussion from the two largest political forums online, and applies textual topic modeling methods to identify which topics the right and left use, and how those topics are connected to each other in a psychological network. This network can then be used to measure more deliberative argumentation, operationalized as making arguments related to those made by one’s interlocutor, rather than merely echoing them or repeating one’s own preferred views. We find interesting differences both between and within the left and right in the degree of deliberativeness; the types of topics preferred and the personality traits these reflect; and the structures of individual psychological networks. This model can also be used to predict how opinions change over months and years in response to arguments made and heard earlier. However, with purely observational data, it is impossible to distinguish prediction from causation, especially with such high-dimensional systems such as text. In order measure high-dimensional free-form text persuasion, in the second part of this project, we use machine learning methods to computationally craft persuasive text and test it experimentally. Thousands of sentences in support of Obamacare are scraped from and parameterized using topic modeling. Three-sentence subsets are then experimentally evaluated for their persuasive effects using Mechanical Turk subjects, and the parameterized space of all three-sentence combinations is optimized using a new machine learning method to experimentally test and iteratively improve the textual treatment until the most persuasive three-sentence combination is found. The topic space also allows to discern which topics were most persuasive and how they interact with each other. This combination of observational and experimentally-crafted textual data allows to model political deliberation and persuasion in its full complexity, with a suite of insights into modern political psychology and behavior.