Argument-Aware RAG Improves Political Fallacy Detection

Summary

A new arXiv study introduces a retrieval-augmented method that uses support and attack relations between arguments to guide document retrieval for political debate fallacy analysis. Tested across 42 retrieval configurations and 14 models on the ElecDeb60to20 benchmark, the approach retrieves from a 15GB political knowledge base and reaches macro-F1 scores of up to 0.864 for detection and 0.725 for classification, outperforming non-retrieval baselines.