Multimodal Argumentative Fallacy Classification in Political Debates

AbstractArgumentative fallacy classification plays a crucial role in improving discourse quality by identifying flawed reasoning that may mislead or manipulate audiences. While traditional approaches have primarily relied on textual analysis, they often overlook paralinguistic cues such as intonation and prosody that are present in speech. In this study, we explore how multimodal analysis, in which we combine textual and audio features, can enhance fallacy classification in political debates. We develop and evaluate text-only, audio-only, and multimodal models using the MM-USED-fallacy dataset to assess the contribution of each modality. Our findings indicate that the multimodal model, which integrates linguistic and acoustic signals, outperforms unimodal systems, underscoring the potential of multimodal approaches in capturing complex argumentative structures.