Leveraging Declarative Knowledge in Text and First-Order Logic for Fine-Grained Propaganda Detection