An Entity-Aware Approach to Logical Fallacy Detection in Kremlin Social Media Content

Abstract

Logical fallacy detection has emerged as a novel and challenging task for language models, more complex than traditional fake news or hate speech detection. This research-in-progress examines an Entity-Aware Approach for logical fallacy detection adapted for a timely use case of Kremlin social media content. As part of this study, a curated dataset of tweets about the war in Ukraine published by Russian government accounts, RuFal, is introduced, on which the Entity-Aware Approach is tested. Preliminary results show the Entity-Aware Approach outperforms baseline pre-trained language models by at least 0.83% on the domain non-specific LOGIC dataset and when both directly transferred to and trained on the domain specific RuFal dataset, by at least 3.09% and 0.45%, respectively, showing the Entity-Aware Approach warrants further research.