Adapting Topic Modeling for Computational Analysis of Framing Processes

This thesis investigates a new approach for leveraging hierarchical topic modeling techniques to analyze and compare dominant frames found during major current events. We focus on the COVID-19 pandemic as it was an international crisis at an unprecedented magnitude, and one of the first of its kind to have full media coverage and social media discussion. We present the latent theta role model, a computational approach to framing analysis that develops latent variables in the form of distribution over words and distributions over grammatical relations to help understand the link between words and grammatical relations. With this newfound understanding of topics and theta roles, this technique can provide clearer insights about framing over Latent Dirichlet Allocation (LDA) topic modeling results. As a result, frames can be developed or solidified from previous qualitative framing analysis.