FallacyES-Political: A Multiclass Dataset of Fallacies in Spanish Political Debates

Resumen

Fallacies are pervasive in political discourse, shaping public opinion and influencing decision-making. Automatic detection and classification of fallacies is a challenging task, especially in non-English languages due to limited resources. In this study, we present FallacyES-Political, a novel dataset of fallacies extracted from 19 electoral debates held in Spain over three decades. The dataset comprises nearly 2,000 fallacies categorized into 16 types. To evaluate the dataset’s utility, we conducted a comprehensive benchmarking of state-of-the-art Large Language Models (LLMs) in zero-shot classification. The results highlight the complexity of fallacy classification and the limitations of current LLMs in understanding contextdependent argumentation. Furthermore, we demonstrate the advantages of finetuning a compact, domain-specific model over relying on general-purpose LLMs, achieving notable improvements in classification accuracy with a more sustainable approach.