Introduction
Detecting fake news is a critical challenge, particularly during the "cold start" phase of breaking events. Most state-of-the-art models rely on high-resource signals such as massive training sets, user comments, or social media diffusion patterns. However, when a story first breaks, these signals are absent, leaving only short snippets of text with no engagement history. Standard fine-tuning of Pre-trained Language Models (PLMs) like BERT often fails in these low-resource scenarios because they lack specific context and sufficient labeled examples to generalize effectively.
The HGAPT (Heterogeneous Graph Augmented Prompt Tuning) framework addresses this by leveraging metadata—such as publishers and topics—as structural priors. Even for new stories, the reputation of a publisher or the nature of a topic provides vital clues for veracity detection.
Methodology
HGAPT integrates prompt-based tuning with a graph-based approach to capture relationships between news items via shared metadata.
Model Architecture
The framework treats news detection as a "cloze" task. For instance, a news snippet is formatted as: "News: [Text]. It was [MASK] news." The model then predicts a "verbalizer" token (e.g., "true" or "false") for the mask.
As shown in the architecture, the model employs Joint Learning. It is trained on two simultaneous objectives:
- Node Classification: Correcting the "true/fake" label for each news item using Cross-Entropy (CE) loss.
- Edge Prediction: Learning to predict relationships between news items based on their representations. If two items share a publisher, the model encourages their internal representations to be more similar through a Heterogeneous Graph (HG) loss.
Heterogeneous Graph Construction
Within every training batch, HGAPT dynamically constructs a graph where each news item is a node. Edges are established between nodes that share metadata, such as the same publisher or topic.
This structural approach acts as a regularizer, preventing overfitting on small datasets by forcing the model to recognize that news from the same source should share certain latent characteristics.
Experimental Setup
The researchers evaluated HGAPT on two primary datasets tailored for low-resource settings:
- LiarFewShot: A subset of the LIAR dataset restricted to 512 training samples per class.
- PolitiFactFewShot: A new dataset curated from PolitiFact.com, including metadata like "news source" often missing in other benchmarks.
The model was tested on both 6-class (fine-grained labels from pants-on-fire to true) and 2-class (binary fake vs. true) tasks.
Key Results and Insights
HGAPT consistently outperformed baselines, including standard BERT fine-tuning and other prompt-tuning variants.
- Accuracy: On the LiarFewShot 6-class task, HGAPT achieved 42.6% accuracy, a significant improvement over standard BERT models (33-35%).
- Binary Performance: It reached 73.5% accuracy on LiarFewShot and 76.5% on PolitiFactFewShot.
- Zero-Shot Generalization: When trained on one domain (e.g., Liar) and tested on another (e.g., PolitiFact), HGAPT maintained its lead, suggesting it learns generalizable patterns of misinformation rather than domain-specific keywords.
A key finding was that calculating edge similarity using predicted class vectors was more effective than using raw hidden states (like the [CLS] token). This indicates that the high-level semantic meaning of veracity is a better bridge between related items than literal word embeddings.
Conclusion
HGAPT demonstrates that metadata can be a powerful structural prior in low-resource fake news detection. By combining the linguistic flexibility of prompt tuning with the relational power of heterogeneous graphs, the model effectively mitigates the "cold start" problem. Future work may explore more complex graph structures or incorporating "soft" metadata like sentiment and writing style.
Reference: Ding, X., Li, F., Teng, C., & Ji, D. (2023). HGAPT: Heterogeneous Graph Augmented Prompt Tuning for Low-Resource Fake News Detection. SEKE 2023.
Summary updated