Enhancing text-centric fake news detection via external knowledge distillation from LLMs

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Computational Social Science

Background:

• Fake news detection is crucial for societal well-being, with text content being a primary indicator.
• Existing methods like SLM-based, external knowledge-enhanced, and LLM-based approaches have limitations in generalization, computational cost, and factual reasoning.

Purpose of the Study:

• To propose LEKD, a hybrid approach that overcomes the limitations of current fake news detection methods.
• To enhance the accuracy and efficiency of text-centric fake news detection by synergizing SLMs, external knowledge, and LLMs.

Main Methods:

• LEKD utilizes LLMs to generate external knowledge for training data augmentation.
• A graph-based semantic-aware feature alignment module resolves knowledge contradictions.
• An information bottleneck-based knowledge distillation module ensures implicit feature generation during inference.

Main Results:

• LEKD demonstrates superior performance compared to baseline methods in fake news detection.
• The proposed method effectively combines the strengths of SLMs, external knowledge, and LLMs.
• Experimental results validate the advantages of LEKD on two benchmark datasets.

Conclusions:

• LEKD offers a robust and efficient solution for text-centric fake news detection.
• The integration of diverse AI techniques in LEKD significantly improves detection capabilities.
• This approach addresses key challenges in current fake news detection paradigms.