Improving the Strength of Counter-Arguments by the Premise-Generation-Step with LLM

Abstract

This study proposes a framework for generating counter-arguments in debates using large language models (LLMs), consisting of two steps: the generation of five premises supporting an affirmative claim, and the generation of attacking premises to effectively weaken the opponent's argument. We applied this framework to LLMs, using topics and affirmative arguments collected from past middle and high school English debate competitions. The counter-arguments generated with and without the application of this framework were then evaluated by student judges. The results showed that counter-arguments produced through the two-step process clearly articulated attacking premises and demonstrated a certain level of effectiveness. This study reports on the potential of LLMs to generate high-quality counter-arguments and aims to contribute to their practical use in educational applications.