Fine-Grained Sentence-Level Propaganda Detection in News Articles

Abstract

The rapid proliferation of generative AI has heightened concerns about propagandistic content in online news, underscoring the need for robust automatic detection methods. This work studies sentence‑level propaganda detection in two settings: (i) binary classification (propaganda vs. non‑propaganda) and (ii) multi‑class technique classification. Using BERT‑ and RoBERTa‑based encoders with different loss functions, our models achieve competitive performance across 14 propaganda techniques while maintaining strong results on the non‑propaganda class. In our experiments, focal loss does not yield statistically meaningful gains over class‑weighted cross‑entropy. Models trained with BERT and class‑weighted cross‑entropy provide the most balanced technique‑level performance; however, several low‑frequency techniques remain challenging, likely reflecting limited training instances. As future work, we will target these low‑resource techniques with data‑centric interventions such as corpus scaling, class‑balanced sampling, and data augmentation to reduce disparities across classes. By strengthening the accuracy and reliability of sentence‑level propaganda detection and by clearly specifying modeling choices, loss formulations, and evaluation protocols this work aims to improve transparency and replicability in propaganda detection research.