From Argument Components to Graphs: A Multi-Agent Debate with Confidence Gating for Argument Relations

The paper presents a novel multi-agent framework for Argument Relation Identification and Classification (ARIC) that utilizes a Proponent-Opponent-Judge architecture, enhancing training-free approaches in Argument Mining. It introduces a confidence gating mechanism that selectively engages in debate only for uncertain predictions, resulting in the highest Macro F1 score on the UKP Argument Annotated Essays v2 corpus among training-free methods. This framework not only outperforms fine-tuned RoBERTa models but also provides human-readable debate transcripts, improving interpretability in argument classification tasks.