Unearthing the Efficacy of ChatGPT in Argumentation Analysis: Performance, Potentials and Limitations

Abstract

Argumentation analysis is a critical task aimed at argument pair extraction on dialogues, thereby facilitating easier comprehension of the content. While existing works often employ finetuned-based models to address this task, our study explores a prompt-based framework utilizing Large Language Models (LLMs). We introduce two prompt frameworks, MPCoT and ArgumentAgent, designed for complex argumentative dialogue comprehension. MPCoT dynamically selects reasoning paths based on sample diversity and label distribution, enhancing the performance of Chain-of-Thought, while ArgumentAgent decomposes the argumentation analysis task into cyclic iterative steps of thinking and observing, making it well-suited for complex logical dialogues. Extensive evaluations show that our proposed model surpasses conventional prompts methods like Few-shot or Chain-of-Thought.

Ähnliche Arbeiten