Using Large Language Models to Assist Content Generation in Persuasive Speaking

Abstract

Many factors contribute to persuasive speech in debate. These include eye contact, diction, and quality of information. We focus on argumentation style in this study. We separate argumentation styles into two categories: emotion and evidence. We primed two models using OpenAI GPT-3, which can rewrite a statement with increased emotive and evidentiary persuasiveness, respectively. We studied the interaction of 10 expert debaters with this system, comparing a version where participants had no control over prompt data, versus one where users could select the prompt data themselves. Participants found that a combination of the emotive and evidentiary models is most effective in persuasive speeches, leaning slightly towards evidence. We also found that certain types of evidence, such as citing studies, are preferred more than statistics such as costs. Finally we found that the majority preferred the model for which they had selected prompt data themselves, since its results aligned more with their interests.