CoALFake: Collaborative Active Learning with Human-LLM Co-Annotation for Cross-Domain Fake News Detection

Note: This repository accompanies a paper currently under review

🔍 Overview

CoALFake is a framework designed to tackle the challenge of cross-domain fake news detection. Traditional models often suffer from domain overfitting and require large amounts of labeled data. CoALFake introduces a Human-LLM Co-Annotation approach combined with Domain-Aware Active Learning to enhance performance and scalability across varied domains.

Key Contributions:

• LLM-assisted Annotation: Efficient labeling using large language models, with optional human review.
• Domain Embedding Integration: Learns rich, domain-aware representations.
• Active Learning Strategy: Selects high-value samples with diverse domain representation.
• Cross-Domain Generalization: Outperforms baselines across political, health, and other domains.