Dissertation Oral Presentation

Investigating Information Framing with Natural Language Processing

Speaker
Date
Mon May 22nd 2023, 9:00 - 10:15am
Location
Margaret Jacks Hall, Greenberg Room (Room 126)

Abstract: Words can influence both who and what we believe. Consider these headlines, which report on the same vaccine: "Moderna Vaccine Proves 96% Effective in Teens," "Moderna Vaccine Highly Effective in Adolescents, Company Says.'' The first presents the efficacy as an established fact, while the second presents the efficacy as an assertion made by its manufacturer. Word choice regarding how assertions are made also matters: a company's findings (with semi-factive findstrengthens the validity of their assertionswhile a politician’s admission carries negative connotations about their trustworthiness. Prior work studying information framing, or the representation of knowledge claims, has been restricted to a relatively narrow set of framing strategies over small-scale datasets. As a result, we know little about systematic patterns in information framing within real-world contexts, and even less about its downstream effects on people’s beliefs. In my dissertation I combine methods from semantics, pragmatics, and Natural Language Processing to study information framing in a data-driven way. In my first series of studies, I use my tools to discover linguistic strategies for conveying trust and doubt and I expose partisan patterns in how journalists use information framing to cover scientific issues. Next, I quantitatively measure the persuasiveness of framing strategies in social media arguments and reveal an interaction between epistemic commitment and politeness: high commitment is more persuasive when framing scientific evidence, but viewed as face-threatening and less persuasive when used to frame personal opinions. Finally, I conduct behavioral experiments to study the effect of information framing on attitude formation. I show that presenting information as opinions using non-factive verbs (e.g., "Scientists believe climate change is a serious concern") can affect what audiences think is true, and to a greater extent, whether they think a ground truth exists in the first place. Together, this work contributes to our understanding of the role of language in how people form, update, and spread their beliefs, and contributes tools and datasets that may be applied to public interest messaging, conflict resolution and negotiation, and bias and misinformation detection systems.