Skip to content

Pseudo-labeling

Pseudo-labeling is a semi-supervised learning technique that automatically generates labels for unlabeled data, typically using model predictions, external systems (such as large language models), or heuristics. In fake news detection, pseudo-labels can be generated by LLMs and used to augment training data when human annotations are scarce.

Key observations

LLM-generated pseudo-labels have varying quality: While LLMs can generate pseudo-labels at scale without human effort, their individual reliability is often limited. Aggregation, confidence filtering, and integration with other signals (e.g., graph structure) are necessary to leverage pseudo-labels effectively.

Confidence-based filtering improves quality: Selecting only high-confidence pseudo-labels (e.g., top 5% confidence scores) significantly improves downstream detection performance by reducing noise in the training signal.

Structural integration enhances reliability: Pseudo-labels propagated through graph structures can achieve higher quality than isolated predictions by leveraging consensus across related samples. Masking mechanisms prevent over-reliance on individual noisy labels.

Key papers