HQP: A Human-Annotated Dataset for Detecting Online Propaganda

TL;DR

This work presents HQP: a novel dataset for detecting online propaganda with high-quality labels, and shows empirically that state-of-the-art language models fail in detecting online propaganda when trained with weak labels, and highlights the importance of high-quality labels for sensitive NLP tasks such as propaganda detection.

Abstract

Online propaganda poses a severe threat to the integrity of societies. However, existing datasets for detecting online propaganda have a key limitation: they were annotated using weak labels that can be noisy and even incorrect. To address this limitation, our work makes the following contributions: (1) We present HQP: a novel dataset (N = 30,000) for detecting online propaganda with high-quality labels. To the best of our knowledge, HQP is the first large-scale dataset for detecting online propaganda that was created through human annotation. (2) We show empirically that state-of-the-art language models fail in detecting online propaganda when trained with weak labels (AUC: 64.03). In contrast, state-of-the-art language models can accurately detect online propaganda when trained with our high-quality labels (AUC: 92.25), which is an improvement of ~44%. (3) We show that prompt-based learning using a small sample of high-quality labels can still achieve a reasonable performance (AUC: 80.27) while significantly reducing the cost of labeling. (4) We extend HQP to HQP+ to test how well propaganda across different contexts can be detected. Crucially, our work highlights the importance of high-quality labels for sensitive NLP tasks such as propaganda detection.

Figures (8)

• Figure 1: Data collection of candidate posts for HQP.
• Figure 2: Contingency table comparing weak vs. high-quality labels.
• Figure 3: $t$-SNE visualization showing the representations of the $\mathrm{[CLS]}$ tokens for BERTweet fine-tuned on TWEETSPIN labels (left) and HQP (right).
• Figure 4: Results for prompt-based learning for LM-BFF vs. LM-BFF-AT (left: F1, right: AUC, top: absolute performance, bottom: %-improvement over LM-BFF). EN (NN) refers to the elastic net (neural net) classification head. $k'$ refers to the number of examples sampled from each class for both training and validation. Error bars denote the standard errors across 5 runs.
• Figure B.1: Instructions for annotators for the context of Russian propaganda.