Propaganda Detection — Propaganda and Manipulation Detection

Sentiment analysis assigns an overall polarity — positive, negative, or neutral — to text. Propaganda detection instead identifies specific persuasion techniques such as loaded language, bandwagon, or glittering generalities, often at the level of individual spans, so the output is a set of technique labels rather than a single tone.

Yes. Propaganda detection is a supervised task, so it assumes a labelled training set annotated against your chosen technique taxonomy. The source recommends at least around fifty annotated documents, and without labelled examples the model has nothing to learn from.

The taxonomy is the catalogue of techniques to detect — for example loaded language, oversimplification, bandwagon, and glittering generalities. It defines the label set the model learns and predicts, so it must be fixed before data is annotated; otherwise labels are inconsistent and evaluation is unreliable.

Score predictions against labelled data and inspect misclassifications. Because technique labels are usually imbalanced, rely on per-class precision, recall, and F1 rather than overall accuracy alone.