MALTO at FadeIT: A BERT-Based System for Multi-Label Fallacy Detection in Italian Social Media

Fallacies in social media posts are common and can make texts less trustworthy, and more likely to foster biased views. In this work, we study multi-label fallacy classification in Italian social media posts by detecting 20 different fallacy types that can appear in the same post. We address this task using encoder-based language models. In particular, we build a fallacy classifier based on AlBERTo, a BERT-style encoder for Italian, and use a global multi-label threshold to turn the model’s outputs into the final set of fallacy labels. Our results indicate that paraphrase-based augmentation can degrade performance in fallacy detection, likely by introducing label noise, while the choice of loss function substantially affects the trade-off between micro- and macro-averaged performance