Using Masked Language Models to Classify Narrative Structure in Political Texts

• In Event: Poster Session: Posters in Political Methodology

Abstract

This paper develops an original method for classifying the latent structure of a piece of text using a Masked Language Model to classify the similarity of component parts and then scoring documents by the overall ordering of each category. Though there is a wealth of work categorizing text based on its topic and tone, there is little methodology for sorting text by the way it is structured. This is despite the fact that anyone who speaks or writes thinks carefully about the best order in which to convey information and the subsequent impact on an audience. For instance, relating information as a story rather than in a straightforward man ner has been shown to be more persuasive and increase empathy. However, in contexts where time is limited or an audience is more skeptical, it may be best to lead with the main point. Being able to classify the structure of text will allow us to both understand how politicians convey information and the contexts in which this varies. The benefits of the method include that it can be used on short form documents, which are often more difficult to break into discrete parts, and each stage can be verified by an analyst rather than black-boxing what is being classified.

Author

• Alexis Palmer, Dartmouth College